Object Detection on COCO (Box AP, Latency)
45.2Box APConvNeXt-T + FDConv
Evaluation results
| Method | Links | ||
|---|---|---|---|
| ConvNeXt-T + FDConvBackbone=ConvNeXt-T, Detector=Mask R-CNN, Params=51M, FLOPS=263G, Training Schedule=1x2025.03 | 45.2 | — | |
| ConvNeXt-T + KWBackbone=ConvNeXt-T, Detector=Mask R-CNN, Params=52M, FLOPS=262G, Training Schedule=1x2025.03 | 44.8 | — | |
| Swin-T + FDConvBackbone=Swin-T, Detector=Mask R-CNN, Params=51M, FLOPS=268G, Training Schedule=1x2025.03 | 44.5 | — | |
| ConvNeXt-TBackbone=ConvNeXt-T, Detector=Mask R-CNN, Params=48M, FLOPS=262G, Training Schedule=1x2025.03 | 43.4 | — | |
| Swin-TBackbone=Swin-T, Detector=Mask R-CNN, Params=48M, FLOPS=267G, Training Schedule=1x2025.03 | 42.7 | — | |
| Full-precision#bits (W/A)=32/32, Backbone=ResNet-502024.04 | 42 | — | |
| IGQ-ViT (#groups=12)#bits (W/A)=6/6, #groups=12, Backbone=ResNet-502024.04 | 41.3 | — | |
| IGQ-ViT (#groups=8)#bits (W/A)=6/6, #groups=8, Backbone=ResNet-502024.04 | 41.1 | — | |
| PTQ for ViT#bits (W/A)=6/6, Backbone=ResNet-502024.04 | 40.5 | — |