Implementation of our ECCV 2020 paper The Devil is in Classification: A Simple Framework for Long-tail Instance Segmentation
This repo contains code of Simcal, which won the LVIS 2019 challenge. Note that it can achieve much higher tail class performance by simply change the calibration head from 2-layer fc with random initialization (2fc_rand) to 3-layer fc initialized from original model with standard training (3fc_ft), refer to paper for details. But we did not notice this during the challenge submission and used 2fc_rand, so much higher result of tail clasees on test set is expected with SimCal 3fc_ft.
This project is released under the Apache 2.0 license.
- remove and clean redundant and commented codes
- update script for installing with pytorch 1.1.0 to have faster calibration training
- merge mask r-cnn and htc model test file, add htc calibration code, add Props-GT experiment code
Pull requests to improve the codebase or fix bugs are welcome
Simcal is based on mmdetection, Please refer to INSTALL.md for installation and dataset preparation.
Or run the following installation script:
#!/usr/bin/env bash
conda create -n simcal_mmdet python=3.7
source ~/anaconda3/etc/profile.d/conda.sh
conda init bash
conda activate simcal_mmdet
echo "python path"
which python
conda install pytorch==1.2.0 torchvision==0.4.0 cudatoolkit=9.2 -c pytorch
pip install cython==0.29.12 mmcv==0.2.16 matplotlib terminaltables
pip install "git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI"
pip install opencv-python-headless
pip install Pillow==6.1
pip install numpy==1.17.1 --no-deps
git clone https://github.com/twangnh/SimCal
cd SimCal
pip install -v -e .
To also get instance-centric AP results, please do not install official LVIS api or cocoapi with pip, as we have modified it with a local copy in the repository to additionally calculate instance centric bin AP results (i.e., AP1,AP2,AP3,AP4). We may create a pull request to update the official APIs for this purpose later.
For LVIS dataset, please arrange the data as:
SimCal
├── configs
├── data
│ ├── LVIS
│ │ ├── lvis_v0.5_train.json.zip
│ │ ├── lvis_v0.5_val.json.zip
│ │ ├── images
│ │ │ ├── train2017
│ │ │ ├── val2017
note for LVIS images, you can just create a softlink for the val2017 to point to COCO val2017
For COCO-LT (our sampled long-tail version of COCO, refer to paper for details), please download the sampled
annotation file train_coco2017_LT_sampled.json and put it at
data/coco/annotations/
Calibration uses multi-gpu training to perform bi-level proposal sampling, to run calibration on a model, e.g.,
python tools/train.py configs/simcal/calibration/mask_rcnn_r50_fpn_1x_lvis_agnostic.py --use_model 3fc_ft --exp_prefix xxx --gpus 4/8
will use 3fc_ft head as described in the paper and save calibrated head ckpt with exp_prefix
All the calibrated models reported in the paper are released for reproduction and future research:
| Model | Link |
|---|---|
| r50-ag epoch-12 | Googledrive |
| calibrated cls head | Googledrive |
| Model | Link |
|---|---|
| r50 epoch-12 | Googledrive |
| calibrated cls head | Googledrive |
| Model | Link |
|---|---|
| r50-ag-coco-lt epoch-12 | Googledrive |
| calibrated cls head | Googledrive |
| Model | Link |
|---|---|
| htc-x101 epoch-20 | Googledrive |
| calhead-stege0 | Googledrive |
| calhead-stege1 | Googledrive |
| calhead-stege2 | Googledrive |
To evaluate and reproduce the paper result models, please first download the model checkpoints and arrange them as:
SimCal
├── configs
├── work_dirs
|-- htc
| |-- 3fc_ft_stage0.pth
| |-- 3fc_ft_stage1.pth
| |-- 3fc_ft_stage2.pth
| `-- epoch_20.pth
|-- mask_rcnn_r50_fpn_1x_cocolt_agnostic
| |-- 3fc_ft.pth
| `-- epoch_12.pth
|-- mask_rcnn_r50_fpn_1x_lvis_agnostic
| |-- 3fc_ft.pth
| `-- epoch_12.pth
`-- mask_rcnn_r50_fpn_1x_lvis_clswise
|-- 3fc_ft_epoch.pth
|-- 3fc_ft.pth
`-- epoch_12.pth
mrcnn on lvis, paper result:
Test LVIS r50-ag model (use --eval bbox for box result)
./tools/dist_test.sh configs/simcal/calibration/mask_rcnn_r50_fpn_1x_lvis_agnostic.py 8 --cal_head 3fc_ft --out ./temp.pkl --eval segm
bin 0_10 AP: 0.13286122428874017
bin 10_100 AP: 0.23243947868384135
bin 100_1000 AP: 0.20696891455408
bin 1000_* AP: 0.2615438157753328
bAP 0.20845335832549858
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=300 catIds=all] = 0.222
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=300 catIds=all] = 0.354
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=300 catIds=all] = 0.236
Average Precision (AP) @[ IoU=0.50:0.95 | area= s | maxDets=300 catIds=all] = 0.154
Average Precision (AP) @[ IoU=0.50:0.95 | area= m | maxDets=300 catIds=all] = 0.298
Average Precision (AP) @[ IoU=0.50:0.95 | area= l | maxDets=300 catIds=all] = 0.373
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=300 catIds= r] = 0.182
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=300 catIds= c] = 0.215
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=300 catIds= f] = 0.247
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=300 catIds=all] = 0.315
Average Recall (AR) @[ IoU=0.50:0.95 | area= s | maxDets=300 catIds=all] = 0.216
Average Recall (AR) @[ IoU=0.50:0.95 | area= m | maxDets=300 catIds=all] = 0.382
Average Recall (AR) @[ IoU=0.50:0.95 | area= l | maxDets=300 catIds=all] = 0.453
Test LVIS r50 model (use --eval bbox for box result)
./tools/dist_test.sh configs/simcal/calibration/mask_rcnn_r50_fpn_1x_lvis_clswise.py 8 --cal_head 3fc_ft --out ./temp.pkl --eval segm
bin 0_10 AP: 0.10187003036862649
bin 10_100 AP: 0.23907519508889202
bin 100_1000 AP: 0.22468457541750592
bin 1000_* AP: 0.28687985066050825
bAP 0.21312741288388318
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=300 catIds=all] = 0.234
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=300 catIds=all] = 0.375
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=300 catIds=all] = 0.245
Average Precision (AP) @[ IoU=0.50:0.95 | area= s | maxDets=300 catIds=all] = 0.167
Average Precision (AP) @[ IoU=0.50:0.95 | area= m | maxDets=300 catIds=all] = 0.316
Average Precision (AP) @[ IoU=0.50:0.95 | area= l | maxDets=300 catIds=all] = 0.405
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=300 catIds= r] = 0.164
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=300 catIds= c] = 0.225
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=300 catIds= f] = 0.272
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=300 catIds=all] = 0.331
Average Recall (AR) @[ IoU=0.50:0.95 | area= s | maxDets=300 catIds=all] = 0.233
Average Recall (AR) @[ IoU=0.50:0.95 | area= m | maxDets=300 catIds=all] = 0.399
Average Recall (AR) @[ IoU=0.50:0.95 | area= l | maxDets=300 catIds=all] = 0.481
mrcnn on cocolt, paper result:
Test COCO-LT r50-ag model (use --eval bbox for box result)
./tools/dist_test.sh configs/simcal/calibration/mask_rcnn_r50_fpn_1x_lvis_agnostic.py 8 --cal_head 3fc_ft --out ./temp.pkl --eval segm
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.246
bin ins nums: [4, 24, 32, 20]
bins ap: [0.1451797625472811, 0.1796142130031695, 0.27337165679657216, 0.3027201541441131]
eAP : 0.22522144662278398
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.412
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.257
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.133
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.278
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.334
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.239
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.424
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.450
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.269
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.481
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.586
htc on lvis, paper result:
Test HTC model (use --eval bbox for box result)
./tools/dist_test_htc.sh configs/simcal/calibration/htc_lvis_31d9.py 8 --out ./temp2.pkl --eval segm
bin 0_10 AP: 0.18796762487467375
bin 10_100 AP: 0.34907335159564473
bin 100_1000 AP: 0.3304618611020927
bin 1000_* AP: 0.3674197862439286
bAP 0.30873065595408494
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=300 catIds=all] = 0.334
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=300 catIds=all] = 0.490
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=300 catIds=all] = 0.357
Average Precision (AP) @[ IoU=0.50:0.95 | area= s | maxDets=300 catIds=all] = 0.228
Average Precision (AP) @[ IoU=0.50:0.95 | area= m | maxDets=300 catIds=all] = 0.422
Average Precision (AP) @[ IoU=0.50:0.95 | area= l | maxDets=300 catIds=all] = 0.565
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=300 catIds= r] = 0.247
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=300 catIds= c] = 0.337
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=300 catIds= f] = 0.364
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=300 catIds=all] = 0.428
Average Recall (AR) @[ IoU=0.50:0.95 | area= s | maxDets=300 catIds=all] = 0.300
Average Recall (AR) @[ IoU=0.50:0.95 | area= m | maxDets=300 catIds=all] = 0.506
Average Recall (AR) @[ IoU=0.50:0.95 | area= l | maxDets=300 catIds=all] = 0.631
note testing with LVIS can be significantly slower as max_det is 300 and det confidence threshold is 0.0
By Props-GT experiment, we would like to emphasize that there is still large room of improvement along the direction of improving object proposal classification.
We also encourage you to check our following up work Balanced Group Softmax after the LVIS challenge, (accepted by CVPR20 oral). It employs a more specific calibration approach with redesigned the softmax function, the calibration is more effective without dual-head inference, and only calibrates last layer of classification head. Code is available at https://github.com/FishYuLi/BalancedGroupSoftmax
Please consider to cite our ECCV20 paper:
@article{wang2020devil,
title={The Devil is in Classification: A Simple Framework for Long-tail Instance Segmentation},
author={Wang, Tao and Li, Yu and Kang, Bingyi and Li, Junnan and Liew, Junhao and Tang, Sheng and Hoi, Steven and Feng, Jiashi},
journal={arXiv preprint arXiv:2007.11978},
year={2020}
}
tech report for LVIS challenge 2019 at ICCV19 (Yu Li and Tao Wang have equal contribution for the LVIS challenge):
@article{wang2019classification,
title={Classification Calibration for Long-tail Instance Segmentation},
author={Wang, Tao and Li, Yu and Kang, Bingyi and Li, Junnan and Liew, Jun Hao and Tang, Sheng and Hoi, Steven and Feng, Jiashi},
journal={arXiv preprint arXiv:1910.13081},
year={2019}
}
Our following work Group Softmax at CVPR20 (oral):
@inproceedings{li2020overcoming,
title={Overcoming Classifier Imbalance for Long-Tail Object Detection With Balanced Group Softmax},
author={Li, Yu and Wang, Tao and Kang, Bingyi and Tang, Sheng and Wang, Chunfeng and Li, Jintao and Feng, Jiashi},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={10991--11000},
year={2020}
}