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

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... (read more)

PDF Abstract

Results from the Paper


TASK DATASET MODEL METRIC NAME METRIC VALUE GLOBAL RANK RESULT BENCHMARK
Real-Time Object Detection COCO CSPResNeXt50-PANet-SPP MAP 33.4 # 14
FPS 58 # 3
inference time (ms) 17 # 3
Object Detection COCO test-dev CSP-p7 + Mish (single-scale) box AP 55.4 # 2
AP50 73.3 # 3
AP75 60.7 # 2
APS 38.1 # 2
APM 59.5 # 2
APL 67.4 # 3
Object Detection COCO test-dev CSP-p7 + Mish (multi-scale) box AP 55.8 # 1
AP50 73.2 # 4
AP75 61.2 # 1
APS 38.8 # 1
APM 60.1 # 1
APL 68.2 # 1
Image Classification ImageNet CSPResNeXt-50 (Mish+Aug) Top 1 Accuracy 79.8% # 70
Top 5 Accuracy 95.2% # 38
Number of params 20.5M # 48

Methods used in the Paper