CSPNet A New Backbone that can Enhance Learning
CSPNet: A New Backbone that can Enhance Learning Capability of CNN Chien-Yao Wang et. al. Taiwan Van-Thanh Hoang thanhhv@islab. ulsan. ac. kr August 8, 2020
Introduction Propose Cross Stage Partial Network (CSPNet) Reduces computations by 20% with equivalent or even superior accuracy on the Image. Net dataset Significantly outperforms state-of-the-art approaches in terms of AP 50 on the MS COCO object detection dataset Easy to implement and general enough to cope with architectures based on Res. Net, Res. Ne. Xt, and Dense. Net 2
CSPDense. Net 3
CSPDense. Net Partial Dense Block Increase gradient path Balance computation of each layer Reduce memory traffic 4
CSPRes. Ne(X)t 5
Exact Fusion Model 6
Experiments Implementation Details Image. Net for classification with batch size of 128 Coco for object detection with batch size of 64 Implement on Darknet framework Train on single GPU 7
Ablation Experiments Ablation study of CSPNet on Image. Net 8
Ablation Experiments Ablation study of EFM on MS COCO GIo. U can upgrade AP by 0. 7%, the AP 50 is significantly degraded by 2. 7% 9
Image. Net Image Classification 10
MS COCO Object Detection 11
Analysis Computational Bottleneck 12
Analysis Memory Traffic 13
Analysis Inference Rate 14
Conclusion CSPNet Respect the variability of the gradients by integrating feature maps from the beginning and the end of a network stage Lack of detailed information of implementation Comments Good idea for balance between accuracy and computation cost 15
Thank you for your attention! 16
PANet 17
Thunder. Net Overall architecture 18
Thunder. Net 19
Generalized Intersection over Union 20
Spatial Pyramid Pooling (SPP) 21
Swish Activation 22
Res 2 Net 23
- Slides: 23