Light-Head R-CNN: In Defense of Two-Stage Object Detector

By Zeming Li, Chao Peng, Gang Yu, Xiangyu Zhang, Yangdong Deng. arXiv.org, 2017

393 citations recorded by the source corpus.

In this paper, we first investigate why typical two-stage methods are not as fast as single-stage, fast detectors like YOLO and SSD. We find that Faster R-CNN and R-FCN perform an intensive computation after or before RoI warping. Faster R-CNN involves two fully connected layers for RoI recognition, while R-FCN produces a large score maps. Thus, the speed of these networks is slow due to the heavy-head design in the architecture. Even if we significantly reduce the base model, the computation cost cannot be largely decreased accordingly. We propose a new two-stage detector, Light-Head R-CNN, to address the shortcoming in current two-stage approaches. In our design, we make the head of network as light as possible, by using a thin feature map and a cheap R-CNN subnet (pooling and single fully-connected layer). Our ResNet-101 based light-head R-CNN outperforms state-of-art object detectors on COCO while keeping time efficiency. More importantly, simply replacing the backbone with a tiny network (e.g, Xception), our Light-Head R-CNN gets 30.7 mmAP at 102 FPS on COCO, significantly outperforming the single-stage, fast detectors like YOLO and SSD on both speed and accuracy. Code will be made publicly available.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/eb46cf7e2eb7f2fa8eca7ee6e0b18fe96ea26965

How researchers see this paper

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

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

Object Detection in 20 Years: A Survey (2019, neutral): “In recent years, the bottle-neck design has been widely used for designing lightweight networks [50], [133], [146], [147], [148].”
TPH-YOLOv5: Improved YOLOv5 Based on Transformer Prediction Head for Object Detection on Drone-captured Scenarios (2021, neutral): “RetinaNet[29] 11.81 21.37 RefineDet[60] 14.90 28.76 DetNet59[26] 15.26 29.23 Cascade-RCNN[3] 16.09 31.91 FPN[28] 16.51 32.20 Light-RCNN[25] 16.53 32.78 CornetNet[24] 17.41 34.12 RRNet (2019 2 nd )[4] 29.13 55.82 DPNet-ensemble (2019 SOTA) [8] 29.62 54.00 SMPNet (2020 2 nd )[47] 35.98 59.53 DPNetV3 (”
MiniVLM: A Smaller and Faster Vision-Language Model (2020, background): “Region feature is built on top of the object detector, and the detector can be two-stage [38, 8, 24, 55] or one-stage [37, 48, 29, 27, 6, 58, 53, 49].”
Object Detection in Aerial Images: A Large-Scale Benchmark and Challenges (2021, methodology): “As shown in Table 13, Faster RCNN OBB + RoI Transformer achieves an OBB mAP of 73.76 for DOTA-v1.”
Recent Advances in Deep Learning for Object Detection (2019, background): “Li et al. [168] proposed a new framework Light Head R-CNN which signiﬁcantly reduced the number of channels in the ﬁnal feature map (from 1024 to 16) instead of sharing all computation.”
DC-SPP-YOLO: Dense Connection and Spatial Pyramid Pooling Based YOLO for Object Detection (2019, methodology): “[34] adopted the improved Xception network as the backbone network in the proposed Light-head R-CNN object detection approach and got better performance than YOLO and SSD on the COCO datasets.”
Scale-Aware Trident Networks for Object Detection (2019, methodology): “To further improve the efficiency of Faster R-CNN, RFCN [10] constructs position-sensitive score maps through a fully convolutional network to avoid the RoI-wise head network.”
Gliding Vertex on the Horizontal Bounding Box for Multi-Oriented Object Detection (2019, background): “faster R-CNN head by light-head R-CNN [43] head.”
Bottom-Up Object Detection by Grouping Extreme and Center Points (2019, neutral): “2 LH R-CNN [23] ResNet-101 1000× 600 41.”
CSPNet: A New Backbone that can Enhance Learning Capability of CNN (2019, methodology): “For real-time object detection on CPU or mobile GPU, SSD-based Pelee [35], YOLOv3-based PRN [33], and Light-Head RCNN [16]based ThunderNet [25] all receive excellent performance on object detection.”
CornerNet-Lite: Efficient Keypoint based Object Detection (2019, methodology): “R-FCN [7] replaces the expensive fully connected sub-detection network with a fully convolutional network, and Light-Head R-CNN [28] reduces the cost in R-FCN by applying separable convolution to reduce the number of channels in the feature maps before RoI pooling.”
Bounding Box Regression With Uncertainty for Accurate Object Detection (2018, background): “Two-stage Detectors: Although one-stage detection algorithms [36, 41, 30, 51] are efficient, state-of-the-art object detectors are based on two-stage, proposal-driven mechanism [42, 4, 5, 17, 31, 2].”
R3Det: Refined Single-Stage Detector with Feature Refinement for Rotating Object (2019, methodology): “Many high-performance detection methods are proposed today, such as R-FCN [7], FPN [22], Light Head RCNN [19], etc.”
ThunderNet: Towards Real-Time Generic Object Detection on Mobile Devices (2019, methodology): “Although Light-Head R-CNN [14] uses a lightweight detection head, it is still too heavy when coupled with small backbones and induces imbalance between the backbone and the detection part.”
RepPoints: Point Set Representation for Object Detection (2019, background): “The bestperforming object detectors to date generally follow a multistage recognition paradigm [27, 4, 14, 26, 2, 41, 30], and the bounding box representation appears in almost all stages: 1) as pre-defined anchors that serve as hypotheses over the bounding box space; 2) as refined object proposals ”