Implementation of ECCV20 paper: the devil is in classification: a simple framework for long-tail object detection and instance segmentation

Related tags

Deep Learningobject-detectioninstance-segmentation
Overview

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.

License

This project is released under the Apache 2.0 license.

TODO

  • 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

Installation

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.

Dataset preparation

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/

Training (Calibration)

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

Pre-trained models and calibrated heads

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

Test with pretrained models and calibrated heads

mrcnn on lvis, paper result:

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:

mrcnn on lvis 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:

mrcnn 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

Props-GT experiment

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.

Balanced Group Softmax

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

Citation

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}
}
Owner
twang
make things work
twang
GitHub Repository
PyTorch implementation of "Simple and Deep Graph Convolutional Networks"

Simple and Deep Graph Convolutional Networks This repository contains a PyTorch implementation of "Simple and Deep Graph Convolutional Networks".(http

chenm 253 Dec 08, 2022
Security evaluation module with onnx, pytorch, and SecML.

🚀 🐼 🔥 PandaVision Integrate and automate security evaluations with onnx, pytorch, and SecML! Installation Starting the server without Docker If you

Maura Pintor 11 Apr 12, 2022
End-to-End Dense Video Captioning with Parallel Decoding (ICCV 2021)

PDVC Official implementation for End-to-End Dense Video Captioning with Parallel Decoding (ICCV 2021) [paper] [valse论文速递(Chinese)] This repo supports:

Teng Wang 118 Dec 16, 2022
2021-MICCAI-Progressively Normalized Self-Attention Network for Video Polyp Segmentation

2021-MICCAI-Progressively Normalized Self-Attention Network for Video Polyp Segmentation Authors: Ge-Peng Ji*, Yu-Cheng Chou*, Deng-Ping Fan, Geng Che

Ge-Peng Ji (Daniel) 85 Dec 30, 2022
Code for Dual Contrastive Learning for Unsupervised Image-to-Image Translation, NTIRE, CVPRW 2021.

arXiv Dual Contrastive Learning Adversarial Generative Networks (DCLGAN) We provide our PyTorch implementation of DCLGAN, which is a simple yet powerf

119 Dec 04, 2022
MEND: Model Editing Networks using Gradient Decomposition

MEND: Model Editing Networks using Gradient Decomposition Setup Environment This codebase uses Python 3.7.9. Other versions may work as well. Create a

Eric Mitchell 141 Dec 02, 2022
Towhee is a flexible machine learning framework currently focused on computing deep learning embeddings over unstructured data.

Towhee is a flexible machine learning framework currently focused on computing deep learning embeddings over unstructured data.

1.7k Jan 08, 2023
Preparation material for Dropbox interviews

Dropbox-Onsite-Interviews A guide for the Dropbox onsite interview! The Dropbox interview question bank is very small. The bank has been in a Chinese

386 Dec 31, 2022
Using Tensorflow Object Detection API to detect Waymo open dataset

Waymo-2D-Object-Detection Using Tensorflow Object Detection API to detect Waymo open dataset Result CenterNet Training Loss SSD ResNet Training Loss C

76 Dec 12, 2022
Pytorch tutorials for Neural Style transfert

PyTorch Tutorials This tutorial is no longer maintained. Please use the official version: https://pytorch.org/tutorials/advanced/neural_style_tutorial

Alexis David Jacq 135 Jun 26, 2022
Super-BPD: Super Boundary-to-Pixel Direction for Fast Image Segmentation (CVPR 2020)

Super-BPD for Fast Image Segmentation (CVPR 2020) Introduction We propose direction-based super-BPD, an alternative to superpixel, for fast generic im

189 Dec 07, 2022
Springer Link Download Module for Python

♞ pupalink A simple Python module to search and download books from SpringerLink. 🧪 This project is still in an early stage of development. Expect br

Pupa Corp. 18 Nov 21, 2022
PyTorch implementation of MulMON

MulMON This repository contains a PyTorch implementation of the paper: Learning Object-Centric Representations of Multi-object Scenes from Multiple Vi

NanboLi 16 Nov 03, 2022
Anchor Retouching via Model Interaction for Robust Object Detection in Aerial Images

Anchor Retouching via Model Interaction for Robust Object Detection in Aerial Images In this paper, we present an effective Dynamic Enhancement Anchor

13 Dec 09, 2022
Create UIs for prototyping your machine learning model in 3 minutes

Note: We just launched Hosted, where anyone can upload their interface for permanent hosting. Check it out! Welcome to Gradio Quickly create customiza

Gradio 11.7k Jan 07, 2023
This repo holds codes of the ICCV21 paper: Visual Alignment Constraint for Continuous Sign Language Recognition.

VAC_CSLR This repo holds codes of the paper: Visual Alignment Constraint for Continuous Sign Language Recognition.(ICCV 2021) [paper] Prerequisites Th

Yuecong Min 64 Dec 19, 2022
A PyTorch implementation of the architecture of Mask RCNN

EDIT (AS OF 4th NOVEMBER 2019): This implementation has multiple errors and as of the date 4th, November 2019 is insufficient to be utilized as a reso

Sai Himal Allu 975 Dec 30, 2022
Classical OCR DCNN reproduction based on PaddlePaddle framework.

Paddle-SVHN Classical OCR DCNN reproduction based on PaddlePaddle framework. This project reproduces Multi-digit Number Recognition from Street View I

1 Nov 12, 2021
This repo is for segmentation of T2 hyp regions in gliomas.

T2-Hyp-Segmentor This repo is for segmentation of T2 hyp regions in gliomas. By downloading the model from here you can use it to segment your T2w ima

1 Jan 18, 2022
SimpleDepthEstimation - An unified codebase for NN-based monocular depth estimation methods

SimpleDepthEstimation Introduction This is an unified codebase for NN-based monocular depth estimation methods, the framework is based on detectron2 (

8 Dec 13, 2022