CSPNet: A New Backbone that can Enhance Learning Capability of CNN

By Chien-Yao Wang, Hongpeng Liao, I-Hau Yeh, Yueh-Hua Wu, Ping-Yang Chen, Jun-Wei Hsieh. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2019

4,207 citations recorded by the source corpus.

Neural networks have enabled state-of-the-art approaches to achieve incredible results on computer vision tasks such as object detection. However, such success greatly relies on costly computation resources, which hinders people with cheap devices from appreciating the advanced technology. In this paper, we propose Cross Stage Partial Network (CSPNet) to mitigate the problem that previous works require heavy inference computations from the network architecture perspective. We attribute the problem to the duplicate gradient information within network optimization. The proposed networks respect the variability of the gradients by integrating feature maps from the beginning and the end of a network stage, which, in our experiments, reduces computations by 20% with equivalent or even superior accuracy on the ImageNet dataset, and significantly outperforms state-of-the-art approaches in terms of AP50 on the MS COCO object detection dataset. The CSPNet is easy to implement and general enough to cope with architectures based on ResNet, ResNeXt, and DenseNet.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/b771c57b8fc242ba589d449e04e982e0577b1c54

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

The YOLO Framework: A Comprehensive Review of Evolution, Applications, and Benchmarks in Object Detection (2024, neutral): “Before the introduction of YOLOv9 [132], YOLO-NAS [133] developed by Deci AI marked a significant shift in the evolution of the YOLO framework.”
Object Detection in 20 Years: A Survey (2019, neutral): “, the ResNet [178], CSPNet [183], Hourglass [184], and Swin Transformer [44].”
A review on modern defect detection models using DCNNs – Deep convolutional neural networks (2021, background): “For more details on the structure of their cross stage layers, please refer to [84].”
FSCE: Few-Shot Object Detection via Contrastive Proposal Encoding (2021, methodology): “Then, the base detector is transferred to novel data through fine-tuning on a balanced dataset [8] with novel instances and randomly sampled base instances (Dtrain = Dnovel ∪ Dbase).”
YOLOX: Exceeding YOLO Series in 2021 (2021, methodology): “To give a fair comparison, we adopt the exact YOLOv5’s backbone including modiﬁed CSPNet [31], SiLU activation, and the PAN [19] head.”
YOLOv4-5D: An Effective and Efficient Object Detector for Autonomous Driving (2021, methodology): “Unlike YOLOv3 that uses DarkNet53 as the backbone, the YOLOv4 uses the cross stage partial (CSP) [11] structure shown in Fig.”
A Comprehensive Review of YOLO: From YOLOv1 to YOLOv8 and Beyond (2023, methodology): “Inspired by GiraffeDet [99], CSPNet [61], and ELAN [95], the authors designed a neck that can work in real-time called Efﬁcient-RepGFPN.”
Object detection using YOLO: challenges, architectural successors, datasets and applications (2022, methodology): “The ImageNet pretrained model for classification with the underlying dataset as COCO for experimentation is utilized [73].”
A Comprehensive Review of YOLO Architectures in Computer Vision: From YOLOv1 to YOLOv8 and YOLO-NAS (2023, methodology): “Inspired by GiraffeDet [112], CSPNet [69], and ELAN [108], the authors designed a neck that can work in real-time called Efficient-RepGFPN.”
Visible and Thermal Image-Based Trunk Detection with Deep Learning for Forestry Mobile Robotics (2021, background): “YOLOv4 has a base composed by CSPDarknet53 [60], a intermediary part made by SPP [61] and Path Aggregation Network (PAN) [62], and an head corresponding to YOLOv3 [59].”
YOLOv1 to v8: Unveiling Each Variant–A Comprehensive Review of YOLO (2024, background): “Furthermore, a noteworthy hurdle is the computational complexity associated with object detection architectures, demanding considerable computational resources in terms of power, memory, and time [17], [18].”
PP-YOLOE: An evolved version of YOLO (2022, methodology): “CSPNet[27] utilizes cross stage dense connections to lower computation burden without the loss of precision, which is popular among effective object detectors such as YOLOv5[14], YOLOX[6].”
ASF-YOLO: A Novel YOLO Model with Attentional Scale Sequence Fusion for Cell Instance Segmentation (2023, methodology): “Cross Stage Partial [10] Darknet with 53 convolutional layers (CSPDarknet53) [11] was modified from YOLOv4 [12] and designed as the backbone network of YOLOv5 [8], which contains C3 (CSP bottleneck including 3 convolutional layers) and ConvBNSiLU modules.”
YOLOv10: Real-Time End-to-End Object Detection (2024, methodology): “YOLOv4 [2] and YOLOv5 [20] introduce the CSPNet [63] design to replace DarkNet [47], coupled with data augmentation strategies, enhanced PAN, and a greater variety of model scales, etc .”
A Survey of Modern Deep Learning based Object Detection Models (2021, methodology): “They proposed CSPNet [25] which creates different paths for the gradient ﬂow within the network.”