Recurrent Neural Networks for video object detection
There is lots of scientific work about object detection in images. For many applications like for example autonomous driving the actual data on which classification has to be done are videos. This work compares different methods, especially those which use Recurrent Neural Networks to detect objects in videos. We differ between feature-based methods, which feed feature maps of different frames into the recurrent units, box-level methods, which feed bounding boxes with class probabilities into the recurrent units and methods which use flow networks. This study indicates common outcomes of the compared methods like the benefit of including the temporal context into object detection and states conclusions and guidelines for video object detection networks.
I-a Image and Video Object Detection in general
History of image object detection
Before the advent of Deep Learning, image detection was carried out using classic Machine Learning methods like Discriminative Models e.g. Logistic Regression, Support Vector Machines etc. or Generative Models e.g. Gaussian Mixture Models etc. Bag of Visual Words (BOVW) was another technique that was used for the same purpose. The use of Deep Learning methods for image detection was restrained by two factors, first of all the dearth of large amounts of labeled data and secondly, the computational power. In 2009, Jia Deng et. al. presented a dataset named ImageNet containing millions of images labeled into different categories. A breakthrough paper published in the same year by Rajat Raina et. al. stipulated that using GPUs for deep learning can provide the much-needed computational power. Hence, the door to using deep learning for image classification and detection were open and the first landmark paper in this domain was published in 2012 by Alex Krizhevsky et. al.
Types of image object detectors
Image object detectors can be categorized into two types, single stage and 2-stage image object detectors. A two-stage pipeline firstly generates region proposals, which are then classified and refined while a single-stage method, which is often more efficient but less accurate, directly regresses on bounding boxes and classes.
Why is video object detection harder?
Large number of individual frames within the videos
The occurence of motion blur between different frames
The quality of video datasets are often not ideal
The objects to be detected can get partially occuluded
The objects can have unconventional poses
I-B Recurrent Neural Networks in general
Delimitation to non-recurrent Neural Networks
Non-recurrent Neural Networks process on single inputs for example, a single image. Recurrent Neural Networks process on sequences of data, e.g.: multiple video frames. Recurrent Neural Network’s core concept to enable the sequence processing is parameter sharing across different parts of a model. Parameter sharing can be reached by cycles in the architecture. 
Common Types of Recurent Neural Networks
Most of the papers which are described in this work use two common Recurrent Neural Network approaches. The first one are LSTMs, first mentioned in  in 1997. The key of LSTMs are different gates (forget gate, external input gate, output gate) to control the cell state and the hidden state of the LSTM .
Ii Feature-based Video Object Detection
First, we want to introduce feature-based Video Object Detection methods. As defined, for example, in [1-2] feature-based Video Object Detection methods fuse detectors which integrate features from multiple frames into their video detection. In most papers considered in this work this integration is done by recurrent units.
Ii-a Recurrent Multi-frame Single Shot Detector for Video Object Detection 
Broad, Jones and Lee have in  and  the idea to design a multi-frame video object detector by inserting a fusion-layer into a classical Single Shot Detector (SSD) meta-architecture. Based on this idea they research in two main fields: On the one hand, they investigate different fusion methods and on the other hand they try several SSD meta-architectures [20-2.1.1; 20-3.6.1]. They test their approaches on the KITTI dataset  for autonomous driving and their model improves upon a state-of-the-art SSD model by 2.7 % mAP. Finally they evaluate their best approach on Caltech Pedestrians Dataset  and find similar improvements of 5 % in maP.
SSDs in general consist of two main components. First, an so called Feature Extractor, which gets as an input an image and outputs feature-maps. The other component is an Detection head which creates bounding boxes and class probabilities out of the feature maps. [1-3] Broad, Jones and Lee have the idea to insert a fusion layer in between those two components. As fusion techniques they test simple element-wise operations (e.g.: add or max), simple concatenations of features maps and recurrent layers [1-3]. Their experiments show that the recurrent layers perform best, because they add in contrast to the other two method new parameters to learn temporal context to the network. In addition, they observe that recurrent units do not slow down the computational speed significantly (only 5 fps slower then the baseline SSD) [1-4.1]. As a recurrent unit they use GRU because they observe that the results were similar to the results when they use LSTMs but the GRUs are faster to train [1-3.1.1]. Their final architecture is shown in Figure 1.
In addition Broad, Jones and Lee test different types of SSDs as a baseline for their architecture. For all baseline SSDs the mAP was higher in comparison to the non-recurrent models. The mAP increase by 1.4 to 4.4 percent on KITTI dataset. The best mAP is achieved using SqueezeDet+  as a baseline SSD network [1-4.2].
Finally Broad, Jones and Lee use the Caltech Pedestrians dataset to expolore the effect of the number of prior frames and the frame-rate. They compare the single-frame SSD model, with the RMf-SSD using the prior frames and the RMf-SSD model using the frames t-2, t-4 and t-6. The improvement was 3 percentage points higher by using the frames t-2, t-4 and t-6 in comparison to use the prior 3 frames [1-4.4].
There is only a little information on the training of the Recurrent Multi-frame Single Shot Detector. They use the SequeezeDet Training strategies [23-3.3] and a pre-trained version of the baseline SSD. Finally, they use the SequezzeDet fine tuning strategy to train the whole network afterwards [12-2].
Overall Broad, Jones and Lee show three main things. First they show that including the temporal context improves single frame SSD methods. Moreover they show that Recurrent Units are a good approach to add the temporal context and finally they point that long time context (frames t-2, t-4, t-6) seems to be more important then short-term temporal context. Finally they reach an mAP of 83 % on KITTI dataset [1-4.2] and a miss-rate of 29 % on the Caltech Pedestrians dataset [1-4.1].
Ii-B Mobile Video Object Detection with Temporally Aware Feature Maps 
Liu and Zhu have in  the goal to design a video object detection architecture, which can run real-time (15 fps) on low-powered mobile and embedded devices. The key of their method is to combine convolutional layers with convolutional LSTMs. They investigate on the benefit from adding an LSTM into the baseline SSD, different types of Recurrent Layers (LSTM, GRU and bottleneck LSTMs), different dimensions of their bottleneck LSTMs and different LSTM placement strategies (Single LSTM placement and multiple LSTM placement). They test their model on Imagenet VID 2015 dataset  and reached a mAP of 54.4 % while performing on 15 fps.
In the beginning they simply add one LSTM to their baseline SSDs architecture – MobileNet . They observe that adding the LSTM improves the mAP in comparison to their baseline SSD architecture. Moreover, they investigate that the greatest improve is by adding the LSTM after the 13th convolutional Layer of the SSD. [2-4.2]
Afterwards, they compare LSTMs, GRUs and Bottleneck LSTMs by placing them after the 13th convolutional layer. Bottleneck-LSTMs have been designed by Liu and Zhu to increase the efficiency of LSTMs. They use convolutions with ReLU as activation function and feed Bottleneck feature map with less input channel then the original feature map into the LST. They come to the conclusion that Bottleneck-LSTMs are more effective the LSTMs and in the case of a convolutional kernel greater 1×1 even more effective than GRUs while attaining comparable perfomance. [2-3.4; 2-4.2]
In addition to the bottleneck LSTMs Liu and Zhu extend their network with multipliers. They use , and as multiplier to scale the channel dimension of each layer. They find out that the accuracy of the model remains near constant up to . For the other multipliers they use the values: and . [2-3.3; 2-4.2]
As shown in Fig. 2 the final model uses LSTMs after each convolutional Layer, because Liu and Zhu observe that there is a slight performance improve by adding LSTMs after every feature map and nearly no change in computational cost [2-4.2].
Liu and Zhu confirm the observation from  that adding temporal context information with the help recurrent units improves the detection quality of the baseline SSDs. Moreover they show that architectures using recurrent units can still perform in real-time (15 fps) even on mobile devices. In comparison to the Mobilnet-SSD model they improve the mAP by 4.1 percentage and perform faster on Pixel 2 Phone. Overall they reach an 54.4 % on Imagenet VID.
This architecture can be especially compared to the model described in ”Looking Fast and Slow” because both of them are processing on a mobile device and are feature-based architectures. But ”Looking Fast and Slow” differs because it does not treat every frame in the same way and proceeds with different features extractors on different frames.
Ii-C Feature Selective Small Object Detection via Knowledge-based recurrent attentive neural networks 
[I] The aim of this paper is develop a video object detector for the purpose of autonomous driving. The network developed in this paper is termed as Knowledge-Based Recurrent Attentive Neural Network (KB-RANN).
The main contributions of the authors, Kai Yi et. al., are:
An attention mechanism that works like human cognition. The attention mechanism can detect the salient features, which are important for the object detection problem and propagate them forward. This attention mechanism is based on previous work by Ashish Vaswani et. al. .
A domain and intuitive knowledge module which can use the knowledge about traffic signals to produce feature maps.
The KB-RANN model is tested on KITTI and MS COCO17 dataset , achieving 0.813 mAP and 0.578 mAP repectively, on all classes of those datasets. The model performs better then several popular objection detection algorithms like Faster R-CNN , RetinaNet  and SqueezeDet .
The authors successfully compressed and accelerated their proposed model and deployed it to their own self-developed autonomous car.
[III-A] The authors use SqueezeNet for feature extractor because SqueezeNet can provide AlexNet level of accuracy but with 50 times less parameters. Although there are better performing models like VGGNet and ResNet, but these models are computationally expensive. They make a few changes in the SqueezeNet architecture, which include changing the kernel size and also fine-tuning the backend by adding two fire modules at the end. This modified SqueezeNet architecture is termed as, SqueezeNet+.
[III-B] LSTM  is used as a memory mechanism in order to find long term dependencies between different frames and the characteristics of Attention and LSTM are exploited together by fusing them with each other, this fused module is termed as RANN. As saliency feature extraction is applied on the original deep feature maps, it is possible that some of the information is lost. Multiple RANNs are cascaded together into a chain to minimize the effect of feature information loss. Moreover, the output H/textsubscriptt of each LSTM is concatenated with the original feature map and used as an input for the next attention module to refine the saliency feature extraction process.
[III-C] Lasty, The authors also add a domain and intuitive knowledge module to the KB-RANN architecture. It is assumed that traffic signs detection is one of the most important aspect of autonomous driving. The major focus of attention of a driver is at the center of vision and the traffic lights are always located at a bias from that central region of peoples’ gazes. With these assumptions in place, domain knowledge about traffic signals is learned from the data itself by constraining the distribution to a 2D Gaussian function and learning the mean and covariance matrices from the data. A reverse Gaussian distribution is computed from the learned Gaussian distribution. The feature maps obtained from this reverse Gaussian distribution are concatenated with the feature maps from RANN.
[IV-A,B,C]For training and evaluation of the model, the authors use the KB-RANN model and MS COCO17 dataset. They compare the results with other popular object detection models. On the KITTI dataset the KB-RANN achieves the mAP of 0.813, compared to the 0.763 of SqueezeDet, 0.702 of Faster R-CNN and 0.601 of RetinaNet. The frames-per-second (FPS) at which KB-RANN operates are higher then the compared models. In order to signify the accuracy gain from the attention mechanism and the knowledge module, authors also train and evaluate different architectures, namely KB-RCNN in which the attention mechanism is replaced with convulational layers and RANN, which does not have knowledge-based module. KB-RANN achieves better accuracy then KB-RCNN and RANN on KITTI and MS COCO17 dataset, although the FPS achieved by KB-RCNN due to its recurrent nature are higher. Lastly, a KB-RANN model with parameters trained on the KITTI dataset is also tested on BTSD dataset to demonstrate the knowledge transfer capabilities.
Ii-D Looking fast and slow: memory-guided mobile video object detection 
 Mason Liu et. al. aim to, replicate the capability of a human visual system of obtaining the ”gist” of a scene. Using this sparse information, and amalgamating it with the more thorough information, a human visual system can effectively detect objects in its field of vision. In light of this, the main contributions of this paper are:
A memory module, which can fuse the outputs of those feature extractors.
The model is evaluated on the ImageNet VID 2015 dataset , which is concatenated with extra data from the ImageNet DET  and MS COCO17 datasets . The model manages to achieve a mAP of 0.593 at 72.3 FPS on a Pixel3 mobile device.
[3-3.1]Let m be the number of feature extractors. For the implementation of this paper, the authors constraint , a d let f/textsubscript1 be the feature extractor which is computationally more expensive while f/textsubscript2 be the feature extractor which is cheaper. Both feature extractors use MobileNetV2 architecture . f/textsubscript1 uses a depth of 1.4 with input frames resolution while f/textsubscript2 uses a depth of 0.35 with a lower input frames resolution.
[3-3.2] A modified LSTM  is used as a memory mechanism to preserve long term dependencies. The modifications in the LSTM are responsible for making the memory mechanism faster and also better at preserving long term dependencies. For faster memory mechanism, the authors make three modifications. They introduce bottlenecking and add a skip connection between the bottleneck and the output. Lastly, the LSTM states are grouped together and convulations are applied on each group seperately and the resultant states are then concatenated together to obtain the final states. The grouped convulations provide a speed-up. In order to improve the preservation of long term dependencies by the LSTM, the LSTM states are only updated when f/textsubscript1 is run and not the f/textsubscript2, as feature maps obtained from f/textsubscript2 are not of a higher quality compared to f/textsubscript1 and this can result in loss of important state information in the LSTM.
SSD-style  detection is applied on refined feature maps obtained from the LSTM for classification and obtaining bounding boxes.
[3-3.4] The interleaving policy which defines the feature extractor that should be used next is based on reinforcement learning. The state space of the reinforcement learning policy network consists of the LSTM states, , , as well as the differences between the states at different timestamps i.e. , and lastly the action history . The action space has length m and the action a means that the feature extractor f/textsubscripta should be run.
[3-3.5] It is observed by the authors that, despite the employment of the interleaving policy the real-time detection is limited by the execution of the expensive f/textsubscript1 feature extractor. They introduce an asynchronous framework for running the feature extractors f/textsubscript1 and f/textsubscript2 extractors in parallel. During testing, this asynchronous framework provides better results.
[4-4.1; 4-4.2] As mentioned before, the authors use ImageNet VID 2015 dataset for training and evaluation, which has 30 classes in total, along with the addition of extra data from ImageNet DET and MS COCO17, but this extra data is limited to the classes contained within ImageNet VID. All results are reported, using a Pixel 3 mobile device. The results are compared with, baseline single-frame detection model i.e. MobilenetV2-SSDLite (mAP: 0.420, FPS: 14.4), LSTM-based model i.e. MobilenetV2-SSDLite+LSTM (mAP: 0.451, FPS: 14.6) and the state of the art mobile video object detection model of Zhu et. al. (mAP: 0.602, FPS: 25.6). The proposed model by authors manages to achieve a mAP of 0.593 at 72.3 FPS. The authors also perform evaluation using slight variations of the proposed architecture i.e. Non-interleaved, Interleaved only, Interleaved+Async and Interleaved+Adaptive+Async, in order to test the significance of different components of the architecture on the end results. The Interleaved+Adaptive+Async provides the most balanced end result. .
Ii-E Detect to Track and track to detect
 In this paper, Christoph Feichtenhofer, Axel Pinz and Andrew Zisserman, aim to perform, detection and tracking simultaneously using a fully convulational network by infering a ”tracklet” over multiple frames. This paper is based on the R-FCN model and extends it for multiple frame and tracking. The main contributions of this paper are:
Finding correlation maps of two feature maps of adjacent frames. These correlation maps are used to estimate the local displacement between frames.
Using ROI-tracking layer to regress bounding boxes over multiple frames.
Linking the detections based on tracklets to infer long-term tubes of objects over the course of the video.
[3-3.2] The R-FCN detection process consists of two stages, firstly a Region Proposal Network (RPN) is used to find the candidate region of interests (ROIs) and then a ROI pooling layer is used to perform classification of the regions into classes or the background. The input to the ROI pooling layer comes from a convulational layer put in place at the end of the ResNet-101 feature extractor with output . The ROI layer outputs position sensitive feature maps and using a softmax layer these position senstive feature maps can be converted to class probablities for each ROI. On a second branch, R-FCN places a convulational layer at the end of the ResNet-101 feature extractor which outputs and this ouput is again used as an input to ROI pooling layer which generates the bounding boxes.
[3-3.3] The objective function is given hereby:
In short, the first term contains , which is the classification, cross-entropy loss, the second term contains , which is the bounding boxes regression loss and lastly, in the last term, is the tracking loss across two frames. All of the loss terms are normalized.
[3-3.4] For each pair of adjacent frames It, It+, a bounding box regression layer is introduced that performs position senstive ROI pooling on the concatentation of bounding box regression features , , which are also stacked with correlation maps, to perform bounding box transformation regression between frames. The correlation maps between the two frames are obtained by finding correlation between the feature maps of the two frames. Finding correlation on all the features in the feature maps will result in an explosion of dimensionality so the correlation maps are only limited to the local neighbors. Like mentioned before, the correlation maps are stacked with the bounding box features maps.
 In practical uses, using all frames of the video is not the most efficient way of detection and tracking, as a lot of information between adjacent frames is redundant and also due to GPU memory and computational restrictions, only a certain number of frames can be processed in the GPU at the same time. Due to this, the authors employ a technique similar to action localization to obtain an optimal path through a video.
[5-5.1;5-5.2;5-5.3] As mentioned above, the proposed model is trained and evaluated on the ImageNet VID dataset. A comparison is made with the R-FCN detector (mAP: 0.742), which the proposed model is based on, the ILSVRC 2015 (mAP: 0.738) winner and the ILSVRC 2016 (mAP: 0.762) winner. DT performs the best with a mAP of 0.82. A comparison between slight variations of the DT model are also evaluated. Firstly, using different feature extractor backbones: ResNet-50 (mAP: 0.767), ResNet-101 (mAP: 0.80) and Inception-v4 (mAP: 0.821) and secondly using the ResNet-101 backbone but a temporal sampling rate of 10 (mAP: 0.786).
Iii Box-Level-based Video Object Detection
Bounding Boxes and Class probabilities are fed into the network and are refined temporally and/or spatially.
Iii-a Context Matters: Refining Object Detection in Video with Recurrent Neural Networks 
In  Tripathi’s, Lipton’s, Belongie’s and Nguyen’s architecture consists two parts: A pseudo-labeler, which assigns labels to all video frames and a recurrent unit which refines those pseudo-labels by using the contextual information. Moreover, they describe a training strategy for their architecture and compare their approach to other models on the YouTube-Objects dataset (v2.0)  , which consists of the ten categories airplane, bird, boat, car, cow, dog, horse, mbike, train. Their model reaches an mAP of 68.73 percent which improves the strongest image-based baseline for Youtube-Video Objects dataset of 7.1% [4-Abstract].
The final architecture can be found in Fig 10. Tripathi, Lipton and Belongie first train a pretrained YOLO object detection network  as a pseudo-labeler on the YouTube-Video Dataset. As specified in YOLO they minimize the weighted squared detection loss and optimize classification and localization error simultaneously [4-3].
For training the whole network they use the following loss function to take both accuracy at the target frame and consistency of predictions across adjacent time steps into consideration. They choose the values of , and based on the detection performance on the validation set [4-2.1]:
Tripathi, Lipton, Belongie and Nguygen use the object detection loss () as described in YOLO [26, 4-2.1.1]. The similarity loss () considers the dissimilarity between the pseudo-labels and prediction at each frame t [4-2.1.2], the category loss () takes wrong class probabilities into consideration [4-2.1.3] and the prediction-consistency loss () regularizes the model by encouraging smoothness predictions across the time-steps [4-2.1.4].
During the evaluation they find two possible areas of improvement for their approach. On the one hand the RNN is not able to recover from wrong predictions made by the pseudo-labeler after they have been fed into the RNN. This is a general disadvantage of box-level methods in comparison to feature-level methods. On the other hand, they observe that their network is not robust to motion [4-3.4].
Tripathi, Lipton, Belongie and Nguyen test their model on the Youtube-Objects dataset. Overall they outperform the best no-recurrent architecture (DA Yolo) in their comparison by 7% mAP [4-3.1]. This again confirms the observatoin that adding recurrency to integrate temporal context improves the detection quality.
The architecture mentioned in ”Context Matters: Refining Object Detection in Video with Recurrent Neural Networks” is pretty similar to the one used in . Both are using the YOLO network architecture as a baseline and feed its output into an Recurrent Unit. Main Differences is that  feeds the bounding boxes and in addition some visual features into the recurrent unit.
Iii-B Optimizing Video Object Detection via a Scale-Time Lattice
 The aim of this paper, is to propose an architecture which is balanced and flexible enough to allow prioritisation of accuracy or performance with minimal effort. The primary contributions of this paper published by Kai Chen et. al. are:
The Scale-Time Lattice, which provides a rich design space.
A novel key-frame extraction policy, which is based on the ease of detection.
 The Scale-Time Lattice allows coarse detection, both temporally and spatially and then use temporal propagation and spatial refinement to go from coarse to fine detection.
[4-4.1] The Scale-time lattice is composed of structures which connect together to perform the temporal and spatial operations. These structures are called Propagation and Refinement Units (PRUs). PRUs work on the basis of two operators and , and by carefully allocating resources to the two operators, a balance between the detection accuracy and the performance can be acheived. Assuming that two frames are sampled from a video at time and , both at scale , the operator tries to model the movement of the bounding boxes from time to and from time to irrespective of the scale offset between the scale and the ground truth frame scale . The modeling of the bounding boxes offset from scale to is the task of the operator. The figure depicting the PRU gives more information on this.
[4-4.1] The objective function is given hereby:
Here, N is the number of bounding boxes in the batch, while and are the network output of and . The and are losses of temporal propagation and spatial refinement network, respectively.
[4-4.2] The most straight-forward keyframe extraction policy is a uniform i.e. to use keyframes from the video after uniform intervals but an intelligent keyframe extraction policy can be used to increase the accuracy of the results. To that effect, the authors introduce a ease of detection cofficient e. If during detection, the value of e falls beneath a certain threshold that the sample rate of the frames from the video is increased. This can happen if there are a lot of objects on the screen or the objects are moving too quickly.
[5-5.2] The proposed model has an mAP of 0.79 at 60 FPS on the ImageNet VID dataset.
Iii-C Spatially Supervised Recurrent Convolutional Neural Networks for Visual Object Tracking 
Ning, Zhang, Huang, He, Ren and Wang design in  a combination of box-level and feature-level based Video detectors. They use the YOLO network to create high-level visual features and preliminary location inferences and feed both into a recurrent unit. They test their approach on the OTB-30 dataset  and compared it with 9 different state-of-the-art trackers and outperform all of them. Their architecture, called ROLO, is shown in figure 11. [5-Abstract; 5-4]
They use three phases to train their model: The pre-training phase of convolutional layers for feature learning, the traditional YOLO training phase for object proposal and the LSTM training phase for object tracking. [5-3]
In the pre-training phase convolutional layers, which create feature maps are pretrained on ImageNet data. Afterwards the YOLO-architecture is adopted as detection module and trained by using the traditional YOLO training phase. Lastly they add LSTMs for tracking. The LSTMs are fed with, the feature representations from the convolutional layers, the Bounding boxes from the detection module and the output states from the LSTM in the previous time-step. For training they use the Mean Squared Error (MSE) [5-3]:
As an alternative, Ning, Zhang, Huang, He, Ren and Wang mention a Heatmap as input for the LSTM. Therefor the prediction and the visual features are concatenated before adding them to the LSTM. This is helpful to visualize the intermediate results. [5-3.3]
Special to this architecture is, that LSTM does regression in two folds. There is a regression within one frame between the location inferences and the high-level features and there is also regression over the different frames of the sequence by taking the temporal context into account [5-3.4].
In 2015 Wu, Lim and Yang published a comparison of different online object trackers on their one data set . The best 9 architectures (STRUCK ; CXT ; OAB ; CSK ; VTD ; VTS ; LSK ; TLD ; RS ) were used by Ning, Zhang, Huang, He, Ren and Wang to compare them with their own architecture. They use 30 videos with different objects (e.g.: car, human, Skater, Couple,..) out of the benchmark  dataset. They get the best tracking result in 13 videos and the second best in 3 videos. There is no other tracker that reaches the best results in more then 4 videos.
This approach combines box-level and feature-level methods and reaches good results based on the combination of spatial regression between features and location proposals and the temporal regression in the LSTM. The paper shows a potential way to combine the best of both methods (feature based and box-level approaches) which are shown in the other papers.
Iv Flow-based Object Detection
Iv-B Deep Feature Flow for Video Recognition 
Zhu, Xiong, Dai, Yuan and Wei develop in  a way to use flow nets to detect objects in video frames. As shown in figure 12 their architecture consists of three main parts: A network to create visual features, a network to create class probabilities and bounding boxes respectively the segmentation out of feature maps, and a network which estimates the optical flow. They tested their approach on the Cityscapes dataset  with scenes for autonomous driving and ground-truth for semantic segmentation and the ImageNet VID dataset  with ground truth for object detection.
The paper differs between key frames and other frames. The network to create visual features only processes on key frames. It is a fully convolutional network, which takes as an input the image and outputs a feature maps. As a feature network they use a pretrained version of ResNet . [3-3, 3-4]
For all non-key frames the feature maps are not constructed by the feature network. Instead, they are propagated by a flow network . Zhu, Xiong, Dai, Yuan and Wei use a state-of-the-art CNN based FlowNet architecture  as a flow network. The flow estimation is given by [3-3, 3-4]:
is the key frame’s feature map, is a two dimensional flow field, it projects back the location of an object in the current frame to the location in the key frame by using bilinear interpolation and is scale field [3-3].
Zhu, Xiong, Dai, Yuan and Wei use a fixed key frame scheduling and they see a potential improvement in changing the key frame policy.
Overall Zhu, Xiong, Dai und Yuan show a alternative to recurrent neural networks to integrate the temporal context into the object tracking. They reach an mIoU up to 69.2 percent on Cityspace dataset and an mAP of 73.9 percent on ImageNet Vid. But those results are reached by processing with 5.6 and 4.05 fps. Both drop significantly with higher fps. Some recurrent papers in 2 and 3 reach higher mAPs on higher fps.
V Comparison of different approaches
V-a KITTI Dataset
|Recurrent ||86.0||50||Nvidia TITAN X||Feature-Level|
|Feature Selective ||81.3||30.8||Nvidia TITAN X||Feature-Level|
As seen in Table I the architecture mentioned in ”Recurrent Multi-frame Single Shot Detector for Video Object Detection”  outperforms the results of the paper ”Feature Selective Small Object Detection via Knowledge-based Recurrent Attentive Neural Network”  in regard to the detection quality (mAP) and the computational speed (fps).
The main difference between those two architectures is that  performs on a single frame and uses LSTMs to include the spatial context within this frame.  instead is processing on a sequence of input frames and uses recurrent units to take the temporal context into consideration.
That leads to the hypothesis that performing on multiple frames is more beneficial than performing on only one frame. Which means that temporal context is more important than spatial context.
V-B ImageNet Dataset
|DT ||82.0||7||Nvidia TITAN X||Feature-Level|
|DT ||78.5||55||Nvidia TITAN X||Feature-Level|
|Scale-Time Lattice ||79.6||20||Nvidia TITAN X||Box-Level|
|Scale-Time Lattice ||79||62||Nvidia TITAN X||Box-Level|
|DeepFeature Flow ||73.9||3||–||Flow-Based|
|DeepFeature Flow ||73.1||20.5||–||Flow-Based|
|Looking Fast and Slow ||60.7||48.8||Pixel Phone||Feature-Level|
|Object Detection with Temporally-Aware ||54.4||15||Pixel Phone||Feature-Level|
As seen in Table 2 both ”Detect to Track and Track to Detect” and ”Optimizing Video Detection with spatial-time lattice” reach better results than the other papers. Both ”Detect to Track and track to Detect” and ”Optimizing Video Detection with spatial-time lattice” perform on multiple frames in parallel. That leads to the advantage that both of them can use temporal context from the past and also from the future. The doubling of the temporal context information is probably one reason for their comparatively good performance.
In addition to the advantage of performing on multiple frames in parallel ”Optimizing Video Detection with spatial-time lattice” uses not only the temporal context. In addition to the temporal context, this approach takes also different scales into consideration. This is a possible further reason for the good results.
The Flow-based approach  has comparatively bad results on ImageNet. We only evaluated one flow-based paper, but the comparatively bad performance could be an evidence that flow-based approaches’ benefit in comparison to recurrent approaches does not exist or is very small.
”Looking Fast and Slow: Memory-Guided Mobile Video Object Detection” and ”Mobile Video Object Detection with Temporally-Aware Feature Maps” cannot be compared to the other 3 approaches because they are running on mobile devices, which have pretty less computational power than the GPUs which are used by the other papers.
V-C Results on COCO Dataset
V-D Results on YouTube Dataset
”Context Matters” is the only paper of those which we have done some research on, which uses the YouTube Dataset to test their architecture. Unfortunately the results cannot be directly compared to the other papers. Also Tripathi, Lipton, Belongie and Nguygen only compare their model with non-recurrent ones.
V-E Results on OTB Challenge Dataset
Out of the paper which are mentioned in this paper only ”Spatially Supervised Recurrent Convolutional Neural Networks for Visual Object Tracking” uses the OTB Challenge Dataset to evaluate their results. Unfortunately, they are only doing a comparison with non-recurrent approaches.
The main conclusion of our research is that temporal context matters. All the papers came to the results that their models, which use more than one frame to detect objects outperform the models with similar baseline architecture which only proceed on single frames.
Moreover, we have seen that operating on multiple-frames at the same time is a beneficial approach. It doubles the amount of temporal context information, which leads to higher mAPs.
With respect to the computational speed we noticed that the recurrent units should not be too deep. And in addition working only on some keyframes can be beneficial to increase the speed. Therefor a good key-frame policy is needed.
For detection quality and also for computational speed it is beneficial to work on different scales. This enables us to use recurrency to take the spatial and the temporal context into consideration.
Vi-B Further work
The proposed architecture consists of a region proposal network  that is based on the N-Gram concept in Natural Language Processing. Given a window of N previous frames, the RPN proposes the regions where the object bounding boxes could be detected from within the next frame. The RPN (region proposal network) should be recurrent in nature for detecting the temporal dependencies and it is to be really light weight. The ROIs obtained from RPN are fed into the network to make the predictions. So rather than feeding the whole image, feed only the region proposals made by RPN. Lastly, affine transformations can be performed to the output bounding boxes to overlay them over the image.
Alexander Broad, Michael Jones, Teng-Yok Lee. Recurrent Multi-frame Single Shot Detector for Video Object Detection. 2018.
Mason Liu, Menglong Zhu. Mobile Video Object Detection with Temporally-Aware Feature Maps. 2018.
Xizhou Zhu, Yuwen Xiong, Jifeng Dai, Lu Yuan, Yichen Wei. Deep Feature Flow for Video Recognition. 2017.
Subarna Tripathi, Zachary C. Lipton, Serge Belongie, Truong Nguyen. Context Matters: Refining Object Detection in Video with Recurrent Neural Networks, 2016.
Guanghan Ning, Zhi Zhang, Chen Huang, Zhihai He, Xiaobo Ren, Haohong Wang. Spatially Supervised Recurrent Convolutional Neural Networks for Visual Object Tracking. 2016.
Mason Liu, Menglong Zhu, Marie White, Yinxiao Li, Dmitry Kalenichenko. Looking Fast and Slow: Memory-Guided Mobile Video Object Detection. 2019.
Christoph Feichtenhofer, Axel Pinz, Andrew Zisserman. Detect to Track and Track to Detect. 2017.
Kai Kang, Wanli Ouyang, Hongsheng Li, Xiaogang Wang. Object Detection from Video Tubelets with Convolutional Neural Networks. 2016.
Kai Chen, Jiaqi Wang, Shuo Yang, Xingcheng Zhang, Yuanjun Xiong, Chen Change Loy, Dahua Lin. Optimizing Video Object Detection via a Scale-Time Lattice. 2018.
Alexander Broad, Michael Jones, Teng-Yok Lee. Supplementary Material for Recurrent Multi-frame Single Shot Detector for Video Object Detection. 2018.
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. 2009.
Alex Krizhevsky, Ilya Sutskever, Geoffrey E. Hinton. ImageNet Classification with Deep Convolutional Neural Networks. 2012.
Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. 2016.
Joseph Redmon, Ali Farhadi. YOLOv3: An Incremental Improvement.
Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, Alexander C. Berg. SSD: Single Shot MultiBox Detector. 2016.
Sepp Hochreiter, JÃ¼rgen Schmidhuber. Long short-term memory In: Neural Computation (journal), vol. 9, issue 8, p. 1735-1780. 1997.
J. Dai, Y. Li, K. He, and J. Sun. R-FCN: Object detection via
region-based fully convolutional networks. In NIPS, 2016.
Jonathan Huang, Vivek Rathod, Chen Sun, Menglong Zhu, Anoop Korattikara, Alireza Fathi, Ian Fischer, Zbigniew Wojna, Yang Song, Sergio Guadarrama, et al.. Speed/accuracy trade-offs for modern convolutional object detectors . In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR, 2017.
Bichen Wu, Forrest Iandola, Peter H. Jin, Kurt Keutzer. SequeezeDet: Unified, Small, Low Power Fully Convolutional Neural Networks for Real-Time Object Detection for Autonomous Driving. In CVPR Workshops, 2017.
A. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, H. Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017.
Alexander Prest, Vicky Kalogeiton, Christian Leistner, Javier Civera, Cordelia Schmid, Vittorio Ferrari. Youtube-objects dataset v2.0, 2014. URL calvin.inf.ed.ac.uk/datasets/youtube-objects-datasets. University of Edinbough (CALVIN), INRIA Grenoble (LEAR), ETH Zurich (CALVIN).
Joseph Redmon, Santosh Kumar Divvala, Ross B. Girshink, Ali Farhadi. You only look once: Unified, real-time object detection, in CVPR. 2016.
Mark Everingham, Luc Gool, Christopher K. Williams, John Winn and Andrew Zisserman. The pascal visual object classes (voc) challenge. Int. J. Comput. Vision, 88(2):303-338, June 2010, ISSN 0920-5691
Yi Wu, Jongwoo Lim, Ming-Hsuan Yang. Object tracking benchmark. Pattern Analysis and Machine Intelligence. IEEE Transaction on, 37(9):1834-1848, 2015
Radu Timofte, Karel Zimmermann, Luc Van Gool. Multi-view traffic sign detection, recognition, and 3D localisation. 2009.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit. Attention Is All You Need. 2017.
Tsung-Yi Lin, Michael Maire, Serge Belongi, Lubomir Bourde, Ross Girshick et. al. Microsoft COCO: Common Objects in Context. 2015.
T.-Y. Lin, P. Goyal, R. Girshick, K. He, P. Dollar, Focal loss Â´
for dense object detection. 2017.
B. Wu, F. Iandola, P. H. Jin, K. Keutzer. Squeezedet: Unified, small, low power fully convolutional neural networks for real-time object detection for autonomous driving. 2016.
O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al. Imagenet large scale visual recognition challenge. 2015.
M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, L. Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. 2018.
X. Zhu, J. Dai, X. Zhu, Y. Wei, and L. Yuan. Towards high performance video object detection for mobiles. 2018.
S. Ren, K. He, R. Girshick, and J. Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. 2016.
J. Dai, Y. Li, K. He, and J. Sun. R-FCN: Object detection via region-based fully convolutional networks. 2016.
K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. 2016.
S. Saha, G. Singh, M. Sapienza, P. H. Torr, and F. Cuzzolin. Deep learning for detecting multiple space-time action tubes in videos. 2016.
K. Kang, H. Li, J. Yan, X. Zeng, B. Yang, T. Xiao, C. Zhang, Z. Wang, R. Wang, X. Wang, and W. Ouyang. T-CNN: tubelets with convolutional neural networks for object detection from videos. 2016.
J. Yang, H. Shuai, Z. Yu, R. Fan, Q. Ma, Q. Liu, and J. Deng. ILSVRC2016 object detection from video: Team NUIST. 2016.
Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke. Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning. 2016.
Karen Simonyan, Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition. 2015.
Rajat Raina, Anand Madhavan, Andrew Y. Ng. Large-scale Deep Unsupervised Learning using Graphics Processors. 2009.
S. Hare, S. Golodetz, A. Saffari, V. Vineet, M. M. Cheng, S. Hicks, P. Torr. Struck: Structured output tracking with kernels. IEEE Transactions on Pattern Analysis and Machine Intelligence, PP(99):1-1, 2105
Thang Ba Dinh, Nam Vo, GÃ©rard Medioni. Context tracker: Exploring supporters and distracters in unconstrained environments, In CVPR, pages 1177-1184. IEEE, 2011.
Helmut Grabner, Michael Grabner, Horst Bischof. Real-time tracking via on-line boosting. In BMVC volume 1, page 6, 2006.
Joao F Henriques, Rui Cuisero, Pedro Martins, Jorge Batista. Exploiting the circulant structure of tracking-by-detection with kernels. In ECCV, pages 702-715. Springer 2012.
Junseok Kwon, Kyoung Mu Lee. Visual tracking decomposition. In CVPR, pages 1269-1276. IEEE, 2010.
Junseok Kwon, Kyoung Mu Lee. Tracking by sampling trackers. In ICCV, pages 1195-1202. IEEE, 2011.
Zdenek Kalal, Juri Matas, Krystian Mikolajcyk. Pn learning: Bootstrapping binary classifiers by structural constrains. In CVPR, pages 49-56. IEEE, 2010.
Robert T Collins, Yanxi Liu, Marius Leordeanu. Online selection of discriminative tracking features. Pattern Analysis and Machine Intelligence, IEEE Transactions on 27(10): 1631-1643, 2005.
M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benenson, U. Franke, S. Roth, B. Schiele. The cityspaces dataset for semantic urban scene understanding. In CVPR, 2016.
A. Dosovitskiy, P. Fischer, E. Ilg, P. Hauser, C. Hazirbas, V. Golkov. Flownet: Learning optical flow with convolutional networks. In ICCV, 2015.