Image-Level or Object-Level? A Tale of Two Resampling Strategies for Long-Tailed Detection

Figure 1: Our proposed Resampling at image-level and obect-level (RIO).

Project page | Paper

Image-Level or Object-Level? A Tale of Two Resampling Strategies for Long-Tailed Detection.
Nadine Chang, Zhiding Yu, Yu-Xiong Wang, Anima Anandkumar, Sanja Fidler, Jose M. Alvarez.
ICML 2021.

This repository contains the official Pytorch implementation of training & evaluation code and the pretrained models for RIO.

Abstract

Training on datasets with long-tailed distributions has been challenging for major recognition tasks such as classification and detection. To deal with this challenge, image resampling is typically introduced as a simple but effective approach. However, we observe that long-tailed detection differs from classification since multiple classes may be present in one image. As a result, image resampling alone is not enough to yield a sufficiently balanced distribution at the object level. We address object-level resampling by introducing an object-centric memory replay strategy based on dynamic, episodic memory banks. Our proposed strategy has two benefits: 1) convenient object-level resampling without significant extra computation, and 2) implicit feature-level augmentation from model updates. We show that image-level and object-level resamplings are both important, and thus unify them with a joint resampling strategy (RIO). Our method outperforms state-of-the-art long-tailed detection and segmentation methods on LVIS v0.5 across various backbones.

Requirements

• Linux or maxOS with Python >= 3.6
• PyTorch >= 1.5 and torchvision corresponding to PyTorch installation. Please refer to download guildlines at the PyTorch website
• Detectron2
• OpenCV is optional but required for visualizations

Installation

Detectron2

Please refer to the installation instructions in Detectron2.

We use Detectron2 v0.3 as the codebase. Thus, we advise installing Detectron2 from a clone of this repository.

LVIS Dataset

Dataset download is available at the official LVIS website. Please follow Detectron's guildlines on expected LVIS dataset structure.

Our Setup

• Python 3.6.9
• PyTorch 1.5.0 with CUDA 10.2
• Detectron2 built from this repository.

Pretrained Models

Detection and Instance Segmentation on LVIS v0.5

Training & Evaluation

Our code is located under projects/RIO.

Our training and evaluation follows those of Detectron2's. We've provided config files for both LVISv0.5 and LVISv1.0.

Example: Training LVISv0.5 on Mask-RCNN ResNet-50

# We advise multi-gpu training
cd projects/RIO
python memory_train_net.py \
--num-gpus 4 \
--config-file=configs/LVISv0.5-InstanceSegmentation/memory_mask_rcnn_R_50_FPN_1x.yaml 

Example: Evaluating LVISv0.5 on Mask-RCNN ResNet-50

cd projects/RIO
python memory_train_net.py \
--eval-only MODEL.WEIGHTS /path/to/model_checkpoint \
--config-file configs/LVISv0.5-InstanceSegmentation/memory_mask_rcnn_R_50_FPN_1x.yaml  

By default, LVIS evaluation follows immediately after training.

Visualization

Detectron2 has built-in visualization tools. Under tools folder, visualize_json_results.py can be used to visualize the json instance detection/segmentation results given by LVISEvaluator.

python visualize_json_results.py --input x.json --output dir/ --dataset lvis

Further information can be found on Detectron2 tools' README.

License

Please check the LICENSE file. RIO may be used non-commercially, meaning for research or evaluation purposes only. For business inquiries, please contact [email protected].

Citation

@article{chang2021image,
  title={Image-Level or Object-Level? A Tale of Two Resampling Strategies for Long-Tailed Detection},
  author={Chang, Nadine and Yu, Zhiding and Wang, Yu-Xiong and Anandkumar, Anima and Fidler, Sanja and Alvarez, Jose M},
  journal={arXiv preprint arXiv:2104.05702},
  year={2021}
}