Please refer to README.md for installation and dataset preparation instructions.
| name | Pretrained Model | Method | Lr Schd | box mAP | mask mAP | #params | FLOPs | Fine-tuned Model |
|---|---|---|---|---|---|---|---|---|
| HorNet-T (7x7) | Tsinghua Cloud | Cascade Mask R-CNN | 3x | 51.7 | 44.8 | 80M | 730G | Tsinghua Cloud |
| HorNet-T (GF) | Tsinghua Cloud | Cascade Mask R-CNN | 3x | 52.4 | 45.6 | 80M | 728G | Tsinghua Cloud |
| HorNet-S (7x7) | Tsinghua Cloud | Cascade Mask R-CNN | 3x | 52.7 | 45.6 | 107M | 830G | Tsinghua Cloud |
| HorNet-S (GF) | Tsinghua Cloud | Cascade Mask R-CNN | 3x | 53.3 | 46.3 | 108M | 827G | Tsinghua Cloud |
| HorNet-B (7x7) | Tsinghua Cloud | Cascade Mask R-CNN | 3x | 53.3 | 46.1 | 144M | 969G | Tsinghua Cloud |
| HorNet-B (GF) | Tsinghua Cloud | Cascade Mask R-CNN | 3x | 54.0 | 46.9 | 146M | 965G | Tsinghua Cloud |
| HorNet-L (7x7) | Tsinghua Cloud | Cascade Mask R-CNN | 3x | 55.4 | 48.0 | 251M | 1363G | Tsinghua Cloud |
| HorNet-L (GF) | Tsinghua Cloud | Cascade Mask R-CNN | 3x | 56.0 | 48.6 | 259M | 1358G | Tsinghua Cloud |
To train a detector with pre-trained models, run:
# multi-gpu training
bash dist_train.sh <CONFIG_FILE> <GPU_NUM> --cfg-options model.pretrained=<PRETRAIN_MODEL> [other optional arguments]
For example, to train a Cascade Mask R-CNN model with a HorNet-T (GF) backbone and 8 gpus, run:
bash dist_train.sh configs/hornet/cascade_mask_rcnn_hornet_tiny_gf_3x_coco_in1k.py 8 --cfg-options model.pretrained=/path/to/pretrained
More config files can be found at configs/hornet.
# multi-gpu testing
bash dist_test.sh <CONFIG_FILE> <DET_CHECKPOINT_FILE> <GPU_NUM> --eval bbox segm
This code is built using mmdetection, timm libraries, and ConvNeXt