This is the implementation of "CSPNet: A New Backbone that can Enhance Learning Capability of CNN" using Pytorch framwork.
For installing Pytorch YOLOv3, you can refer to YOLOv3(ultralytics).
This branch shows the results train CSPNet from scratch using Pytorch.
| Model | Size | NMS | 1080ti fps | BFLOPs | AP | AP50 | AP75 | cfg | weight |
|---|---|---|---|---|---|---|---|---|---|
| YOLOv3-SPP (baseline) | 512×512 | 0.5 | 50 | 100.343 | 39.7 | 60.5 | 42.2 | cfg | weight |
| CSPResNeXt50c-YOLO-SPP | 512×512 | 0.5 | 43 | 58.983 | 38.4 | 59.6 | 40.5 | cfg | weight |
| CSPDarknet53-YOLO-SPP | 512×512 | 0.5 | - | 75.513 | 39.2 | 60.2 | 41.6 | cfg | - |
| CSPResNeXt50-PANet-SPP | 512×512 | 0.5 | 44 | 71.331 | 39.2 | 59.5 | 41.8 | cfg | - |
| CSPResNeXt50c-PANet-SPP | 512×512 | 0.5 | - | 71.734 | 39.9 | 60.1 | 42.6 | cfg | - |
| CSPResNet50c-PANet-SPP | 512×512 | 0.5 | 56 | 74.955 | 38.4 | 58.5 | 41.0 | cfg | weight |
| CSPDarknet53s-PANet-SPP | 512×512 | 0.5 | 50 | 88.398 | 40.0 | 60.4 | 42.9 | cfg | weight |
| CSPDarknet53m-PANet-SPP | 512×512 | 0.5 | 48 | 88.264 | 39.8 | 60.1 | 42.6 | cfg | weight |
| YOLOv3-SPP (baseline) | 608×608 | 0.5 | 35 | 141.500 | 40.1 | 60.9 | 42.8 | - | - |
| CSPResNeXt50c-YOLO-SPP | 608×608 | 0.5 | 34 | 83.176 | 38.9 | 60.3 | 41.3 | - | - |
| CSPDarknet53-YOLO-SPP | 608×608 | 0.5 | - | 106.485 | 39.6 | 60.7 | 42.3 | - | - |
| CSPResNeXt50-PANet-SPP | 608×608 | 0.5 | 35 | 100.588 | 39.7 | 60.2 | 42.4 | - | - |
| CSPResNeXt50c-PANet-SPP | 608×608 | 0.5 | - | 101.156 | 40.2 | 60.5 | 43.1 | - | - |
| CSPResNet50c-PANet-SPP | 608×608 | 0.5 | 40 | 105.699 | 38.9 | 59.2 | 41.6 | - | - |
| CSPDarknet53s-PANet-SPP | 608×608 | 0.5 | 38 | 124.655 | 40.2 | 60.6 | 43.3 | - | - |
| CSPDarknet53m-PANet-SPP | 608×608 | 0.5 | 36 | 124.466 | 40.1 | 60.6 | 43.1 | - | - |
| Model | Size | NMS | 1080ti fps | AP | AP50 | AP75 | cfg | weight |
|---|---|---|---|---|---|---|---|---|
| CSPNet-PANet-SPP | 320×320 | 0.5 | - | 23.8 | 40.5 | 24.2 | - | - |
| CSPNet-YOLOv3-SPP | 320×320 | 0.5 | - | 22.2 | 39.5 | 22.0 | - | - |
| YOLOv3-tiny (baseline) | 416×416 | 0.5 | 330 | 16.6 | 33.0 | 14.9 | - | - |
| CSPNet-PANet-SPP | 416×416 | 0.5 | 238 | 26.5 | 44.8 | 27.0 | - | - |
| CSPNet-PANet-SPP (darknet) | 416×416 | 0.5 | 238 | 24.4 | 45.9 | 23.7 | - | - |
| CSPNet-YOLOv3-SPP | 416×416 | 0.5 | 220 | 24.9 | 43.6 | 24.9 | - | - |
| Model | Size | 1080ti fps | TX2 fps | TX2(TRT-F) fps | Xavier fps | Xavier(TRT-I) fps | AP | AP50 | AP75 |
|---|---|---|---|---|---|---|---|---|---|
| CSPNet-PANet-SPP | 416×416 | 238 | 38 | 44 | 117 | 254 | 26.5 | 44.8 | 27.0 |
※ the current fps is a rough estimation, while i am training other models when testing it.
※ multi-scale training use input size 288 to 640, except CSPResNeXt50c-PANet-SPP use 320 to 608.
@inproceedings{wang2020cspnet,
title={CSPNet: A new backbone that can enhance learning capability of cnn},
author={Wang, Chien-Yao and Mark Liao, Hong-Yuan and Wu, Yueh-Hua and Chen, Ping-Yang and Hsieh, Jun-Wei and Yeh, I-Hau},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops},
pages={390--391},
year={2020}
}