This repo is the official implementation for CVPR 2021 paper: Adaptive Class Suppression Loss for Long-Tail Object Detection.


1. Environment:

The requirements are exactly the same as BalancedGroupSoftmax. We tested on the following settings:

  • python 3.7
  • cuda 10.0
  • pytorch 1.2.0
  • torchvision 0.4.0
  • mmcv 0.2.14
conda create -n mmdet python=3.7 -y
conda activate mmdet

pip install cython
pip install numpy
pip install torch
pip install torchvision
pip install pycocotools
pip install matplotlib
pip install terminaltables

# download the source code of mmcv 0.2.14 from
cd mmcv-0.2.14
pip install -v -e .
cd ../

git clone

cd ACSL/lvis-api/
python develop

cd ../
python develop

2. Data:

a. For dataset images:

# Make sure you are in dir ACSL

mkdir data
cd data
mkdir lvis
mkdir pretrained_models
mkdir download_models
  • If you already have COCO2017 dataset, it will be great. Link train2017 and val2017 folders under folder lvis.
  • If you do not have COCO2017 dataset, please download:
    COCO train set and
    COCO val set
    and unzip these files and mv them under folder lvis.

b. For dataset annotations:

c. For pretrained models:

Download the corresponding pre-trained models below.

  • To train baseline models, we need models trained on COCO to initialize. Please download the corresponding COCO models at mmdetection model zoo.

  • Move these model files to ./data/pretrained_models/

d. For download_models:

Download the trained baseline models and ACSL models from BaiduYun, code is 2jp3

  • To train ACSL models, we need corresponding baseline models trained on LVIS to initialize and fix all parameters except for the last FC layer.

  • Move these model files to ./data/download_models/

After all these operations, the folder data should be like this:

    ├── lvis
    │   ├── lvis_v0.5_train.json
    │   ├── lvis_v0.5_val.json
    │   ├── train2017
    │   │   ├── 000000100582.jpg
    │   │   ├── 000000102411.jpg
    │   │   ├── ......
    │   └── val2017
    │       ├── 000000062808.jpg
    │       ├── 000000119038.jpg
    │       ├── ......
    └── pretrained_models
    │       ├── faster_rcnn_r50_fpn_2x_20181010-443129e1.pth
    │       ├── ......
    └── download_models
            ├── R50-baseline.pth
            ├── ......


Note: Please make sure that you have prepared the pretrained_models and the download_models and they have been put to the path specified in ${CONIFG_FILE}.

Use the following commands to train a model.

# Single GPU
python tools/ ${CONFIG_FILE}

# Multi GPU distributed training
./tools/ ${CONFIG_FILE} ${GPU_NUM} [optional arguments]

All config files are under ./configs/.

  • ./configs/baselines: all baseline models.
  • ./configs/acsl: models for ACSL models.

For example, to train a ACSL model with Faster R-CNN R50-FPN:

# Single GPU
python tools/ configs/acsl/

# Multi GPU distributed training (for 8 gpus)
./tools/ configs/acsl/ 8

Important: The default learning rate in config files is for 8 GPUs and 2 img/gpu (batch size = 8*2 = 16). According to the Linear Scaling Rule, you need to set the learning rate proportional to the batch size if you use different GPUs or images per GPU, e.g., lr=0.01 for 4 GPUs * 2 img/gpu and lr=0.08 for 16 GPUs * 4 img/gpu. (Cited from mmdetection.)


Use the following commands to test a trained model.

# single gpu test
python tools/ \

# multi-gpu testing
./tools/ \
  • $RESULT_FILE: Filename of the output results in pickle format. If not specified, the results will not be saved to a file.
  • $EVAL_METRICS: Items to be evaluated on the results. bbox for bounding box evaluation only. bbox segm for bounding box and mask evaluation.

For example (assume that you have finished the training of ACSL models.):

  • To evaluate the trained ACSL model with Faster R-CNN R50-FPN for object detection:
# single-gpu testing
python tools/ configs/acsl/ \
 ./work_dirs/acsl/faster_rcnn_r50_fpn_1x_lvis_tunefc_acsl/epoch_12.pth \
  --out acsl_val_result.pkl --eval bbox

# multi-gpu testing (8 gpus)
./tools/ configs/acsl/ \
./work_dirs/acsl/faster_rcnn_r50_fpn_1x_lvis_tunefc_acsl/epoch_12.pth 8 \
--out acsl_val_result.pkl --eval bbox

Results and models

Please refer to our paper for more details.

Method Models bbox mAP Config file Pretrained Model Model
baseline R50-FPN 21.18 file COCO-R50 R50-baseline
ACSL R50-FPN 26.36 file R50-baseline R50-acsl
baseline R101-FPN 22.36 file COCO-R101 R101-baseline
ACSL R101-FPN 27.49 file R101-baseline R101-acsl
baseline X101-FPN 24.70 file COCO-X101 X101-baseline
ACSL X101-FPN 28.93 file X101-baseline X101-acsl
baseline Cascade-R101 25.14 file COCO-Cas-R101 Cas-R101-baseline
ACSL Cascade-R101 29.71 file Cas-R101-baseline Cas-R101-acsl
baseline Cascade-X101 27.14 file COCO-Cas-X101 Cas-X101-baseline
ACSL Cascade-X101 31.47 file Cas-X101-baseline Cas-X101-acsl

Important: The code of BaiduYun is 2jp3


  title={Adaptive Class Suppression Loss for Long-Tail Object Detection},
  author={Wang, Tong and Zhu, Yousong and Zhao, Chaoyang and Zeng, Wei and Wang, Jinqiao and Tang, Ming},