This is the LabelEnc implementation based on FCOS and maskrcnn-benchmark. Our full paper is available at https://arxiv.org/abs/2007.03282

See INSTALL.md for installation instructions.

Our implementation is based on FCOS and maskrcnn-benchmark. You may refer to them for installation instructions as well.

LabelEnc uses a two-step training pipeline. In step1, we acquire a Label Encoding Function. In step2, we train the final detection model.

You will need to download the COCO dataset. We recommend to symlink the path to the coco dataset to datasets/ as follows

# symlink the coco dataset
cd LabelEnc
mkdir -p datasets/coco
ln -s /path_to_coco_dataset/annotations datasets/coco/annotations
ln -s /path_to_coco_dataset/train2014 datasets/coco/train2014
ln -s /path_to_coco_dataset/test2014 datasets/coco/test2014
ln -s /path_to_coco_dataset/val2014 datasets/coco/val2014

Use the following commands to train LabelEnc Step1 with on 8 GPUs. Take retinanet_R-50-FPN for example:

python -m torch.distributed.launch --nproc_per_node=8 \
    tools/train_labelenc_step1.py \
    --config-file configs/retinanet/retinanet_R-50-FPN_1x.yaml \
    OUTPUT_DIR exps/step1/retinanet_R-50-FPN/

The Label Encoding Function will be saved to exps/step1/retinanet_R-50-FPN/label_encoding_function.pth.

Note that:

1. The shown AP is the precision of the auxiliary detection model in Step1 (denoted in the paper as "Step1 only" in Table 4).
2. We use 1x schedule in both Step1 and Step2. Use only xxx_1x.yaml when training LabelEnc, or change the schedule in the config file to [60000, 80000, 90000] yourself.

Use the following commands to train LabelEnc Step2 with retinanet_R-50-FPN_1x.yaml on 8 GPUs. Use --labelenc command to specify label_encoding_function.pth file. Take retinanet_R-50-FPN for example:

python -m torch.distributed.launch --nproc_per_node=8 \
    tools/train_labelenc_step2.py \
    --config-file configs/retinanet/retinanet_R-50-FPN_1x.yaml \
    --labelenc exps/step1/retinanet_R-50-FPN/label_encoding_function.pth \
    OUTPUT_DIR exps/step2/retinanet_R-50-FPN/

We provide the following trained LabelEnc models.

[1] 1x and 2x only notes the schedule for baseline models. LabelEnc models use 1x in both steps.

[2] We use R_50 models in step1 for all models above, no matter their backbone in step2. For example, retinanet_dcnv2_R-101-FPN uses label_encoding_function.pth from retinanet_R-50-FPN.

Links for step1 weights, i.e. label_encoding_function.pth are listed below.

* We train step1 models individually for fcos_xxx, fcos_imprv_xxx and fcos_imprv_dcnv2_xxx because they have different detection heads (fcos_imprv_dcnv2 uses dcn in head in the official release).

You may cite our work in your publications with the following BibTex codes:

@article{hao2020labelenc,
  title={LabelEnc: A New Intermediate Supervision Method for Object Detection},
  author={Hao, Miao and Liu, Yitao and Zhang, Xiangyu and Sun, Jian},
  journal={arXiv preprint arXiv:2007.03282},
  year={2020}
}