Accurate Single Stage Detector Using Recurrent Rolling Convolution

By Jimmy S. J. Ren, Xiaohao Chen, Jianbo Liu, Wenxiu Sun, Jiahao Pang, Qiong Yan, Yu-Wing Tai, Li Xu. 2017

297 citations recorded by the source corpus.

Most of the recent successful methods in accurate object detection and localization used some variants of R-CNN style two stage Convolutional Neural Networks (CNN) where plausible regions were proposed in the first stage then followed by a second stage for decision refinement. Despite the simplicity of training and the efficiency in deployment, the single stage detection methods have not been as competitive when evaluated in benchmarks consider mAP for high IoU thresholds. In this paper, we proposed a novel single stage end-to-end trainable object detection network to overcome this limitation. We achieved this by introducing Recurrent Rolling Convolution (RRC) architecture over multi-scale feature maps to construct object classifiers and bounding box regressors which are deep in context. We evaluated our method in the challenging KITTI dataset which measures methods under IoU threshold of 0.7. We showed that with RRC, a single reduced VGG-16 based model already significantly outperformed all the previously published results. At the time this paper was written our models ranked the first in KITTI car detection (the hard level), the first in cyclist detection and the second in pedestrian detection. These results were not reached by the previous single stage methods. The code is publicly available.

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

StairNet: Top-Down Semantic Aggregation for Accurate One Shot Detection (2017, background): “RRC [29] transfers each feature semantic information to other layers by stacking pooling and deconvolution layers upon SSD. [8, 25, 26] have shown that encoder-decoder, hourglass shape, is effective to propagate the context information.”
Recent Advances in Deep Learning for Object Detection (2019, background): “Thus, handling feature were later developed [109, 110, 109, 111, 112, 92, 113, 114, 115, 116, 117, 118, 119], with modiﬁcations to the feature pyramid block (see Fig.”
Deep learning in video multi-object tracking: A survey (2019, methodology): “[67] used instead a Recurrent Rolling Convolution (RRC) CNN [68] and a SubCNN [69] to detect vehicles in videos recorded on a moving camera in the context of autonomous driving (see section 3.”
Complex-YOLO: Real-time 3D Object Detection on Point Clouds (2018, background): “s Lidar based only. In general, object detection and classiﬁcation based on deep learning is a well known task and widely established for 2D bounding box regression on images [12] [13] [14] [15] [16] [17] [18] [19] [20] [21]. Research focus was mainly a trade-off between accuracy and efﬁciency. In r”
EuroCity Persons: A Novel Benchmark for Person Detection in Traffic Scenes (2018, methodology): “Apart from [60] all of these are two stage methods building”
Joint 3D Proposal Generation and Object Detection from View Aggregation (2017, background): “The gap between the two remains large on standard benchmarks such as the KITTI Object Detection Benchmark [1] where 2D car detectors such as RRC [2] have achieved over 90% Average Precision (AP), whereas the top scoring 3D car detector on the same data only achieves around 70% 3D AP and 84% bird’s e”
Frustum ConvNet: Sliding Frustums to Aggregate Local Point-Wise Features for Amodal (2019, methodology): “For the KITTI test dataset, we use RRC [26] for the car category and MSCNN [27] for the pedestrians and cyclist categories.”
MOTS: Multi-Object Tracking and Segmentation (2019, methodology): “Instead we used their original detections which are obtained from RRC [48], a very strong detector.”
M3D-RPN: Monocular 3D Region Proposal Network for Object Detection (2019, background): “We observe that 2D object detection performs reasonably and continues to make rapid advances [5, 6, 11, 14, 19, 33].”
Joint Monocular 3D Vehicle Detection and Tracking (2018, methodology): “To make a fair comparison, we use RRC [34] detection results following a published state-of-the-art method, BeyondPixels [42].”
Segment as Points for Efficient Online Multi-Object Tracking and Segmentation (2020, methodology): “For cars, the 3D tracking method MOTSFusion adopts a time-consuming detector RRC [31] which takes 3.”
Simulating Photo-realistic Snow and Fog on Existing Images for Enhanced CNN Training and Evaluation (2019, methodology): “[40], we chose a random subset (468 images which were not used for training) of the KITTI image data set [5] and examined the object detection average precision.”
Cooper: Cooperative Perception for Connected Autonomous Vehicles Based on 3D Point Clouds (2019, background): “Rapid development of autonomous vehicles has motivated research institutions to develop representations to perceive local environment, such as lane detection, traffic sign detection and detect objects like cars, cyclists and pedestrians [19], [20], [29], [31] based on the open datasets [7], [9], [10”
Pseudo-LiDAR From Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving (2018, methodology): “Therefore, for the car category we apply the released RRC detector [10] trained on KITTI (see Table 5 in the main paper).”
Tracking Objects as Points (2020, background): “An off-the-shelf object detector [9, 30, 31, 51] first finds all objects in each individual frame.”
Multi-Task Multi-Sensor Fusion for 3D Object Detection (2019, methodology): “In 2D detection, we surpass the best image detector RRC [19] by 1.1% AP in the hard setting, while being 45× faster.”
Fusing Bird View LIDAR Point Cloud and Front View Camera Image for Deep Object Detection (2017, methodology): “The DNN-based object detection with either LIDAR [2, 3] or camera [4, 5, 6] data as the input has been widely explored by researchers and pushed to a very high single frame accuracy.”