Revisiting Contrastive Methods for Unsupervised Learning of Visual Representations

This repo contains the Pytorch implementation of our paper:

Revisiting Contrastive Methods for UnsupervisedLearning of Visual Representations

Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis and Luc Van Gool.

🆕 Accepted at NeurIPS 2021.

Contents

1. Introduction
2. Key Results
3. Installation
4. Training
5. Evaluation
   • Linear Classifier
   • Semantic Segmentation
   • Segment Retrieval
   • Video Instance Segmentation
6. Model Zoo
7. Citation

Introduction

Contrastive self-supervised learning has outperformed supervised pretraining on many downstream tasks like segmentation and object detection. However, current methods are still primarily applied to curated datasets like ImageNet. We first study how biases in the dataset affect existing methods. Our results show that an approach like MoCo works surprisingly well across: (i) object- versus scene-centric, (ii) uniform versus long-tailed and (iii) general versus domain-specific datasets. Second, given the generality of the approach, we try to realize further gains. We show that learning additional invariances - through the use of multi-scale cropping, stronger augmentations and nearest neighbors - improves the representations. Finally, we observe that MoCo learns spatially structured representations when trained with a multi-crop strategy. The representations can be used for semantic segment retrieval and video instance segmentation without finetuning. Moreover, the results are on par with specialized models. We hope this work will serve as a useful study for other researchers.

Key Results

• Scene-centric Data: We do not observe any indications that contrastive pretraining suffers from using scene-centric image data. This is in contrast to prior belief. Moreover, if the downstream data is non-object-centric, pretraining on scene-centric datasets even outperforms ImageNet pretraining.
• Dense Representations: The multi-scale cropping strategy allows the model to learn spatially structured representations. This questions a recent trend that proposed additional losses at a denser level in the image. The representations can be used for semantic segment retrieval and video instance segmentation without any finetuning.
• Additional Invariances: We impose additional invariances by exploring different data augmentations and nearest neighbors to boost the performance.
• Transfer Performance: We observed that if a model obtains improvements for the downstream classification tasks, the same improvements are not guaranteed for other tasks (e.g. semantic segmentation) and vice versa.

Installation

The Python code runs with recent Pytorch versions, e.g. 1.6. Assuming Anaconda, the most important packages can be installed as:

conda install pytorch=1.6.0 torchvision=0.7.0 cudatoolkit=10.2 -c pytorch
conda install -c conda-forge opencv           # For evaluation
conda install matplotlib scipy scikit-learn   # For evaluation

We refer to the environment.yml file for an overview of the packages we used to reproduce our results. The code was run on 2 Tesla V100 GPUs.

Training

Now, we will pretrain on the COCO dataset. You can download the dataset from the official website. Several scripts in the scripts/ directory are provided. It contains the vanilla MoCo setup and our additional modifications for both 200 epochs and 800 epochs of training. First, modify --output_dir and the dataset location in each script before executing them. Then, run the following command to start the training for 200 epochs:

sh scripts/ours_coco_200ep.sh # Train our model for 200 epochs.

The training currently supports:

• MoCo
• + Multi-scale constrained cropping
• + AutoAugment
• + kNN-loss

A detailed version of the pseudocode can be found in Appendix B.

Evaluation

We perform the evaluation for the following downstream tasks: linear classification (VOC), semantic segmentation (VOC and Cityscapes), semantic segment retrieval and video instance segmentation (DAVIS). More details and results can be found in the main paper and the appendix.

Linear Classifier

The representations can be evaluated under the linear evaluation protocol on PASCAL VOC. Please visit the ./evaluation/voc_svm directory for more information.

Semantic Segmentation

We provide code to evaluate the representations for the semantic segmentation task on the PASCAL VOC and Cityscapes datasets. Please visit the ./evaluation/segmentation directory for more information.

Segment Retrieval

In order to obtain the results from the paper, run the publicly available code with our weights as the initialization of the model. You only need to adapt the amount of clusters, e.g. 5.

Video Instance Segmentation

In order to obtain the results from the paper, run the publicly available code from Jabri et al. with our weights as the initialization of the model.

Model Zoo

Several pretrained models can be downloaded here. For a fair comparison, which takes the training duration into account, we refer to Figure 5 in the paper. More results can be found in Table 4 and Table 9.

Citation

This code is based on the MoCo repository. If you find this repository useful for your research, please consider citing the following paper(s):

@inproceedings{vangansbeke2021revisiting,
  title={Revisiting Contrastive Methods for Unsupervised Learning of Visual Representations},
  author={Van Gansbeke, Wouter and Vandenhende, Simon and Georgoulis, Stamatios and Van Gool, Luc},
  booktitle={Advances in Neural Information Processing Systems},
  year={2021}
}
@inproceedings{he2019moco,
  title={Momentum Contrast for Unsupervised Visual Representation Learning},
  author={Kaiming He and Haoqi Fan and Yuxin Wu and Saining Xie and Ross Girshick},
  booktitle = {Conference on Computer Vision and Pattern Recognition},
  year={2019}
}

For any enquiries, please contact the main authors.

Extra

• For an overview on self-supervised learning (SSL), have a look at the overview repository.
• Interested in self-supervised semantic segmentation? Check out our recent work: MaskContrast.
• Interested in self-supervised classification? Check out SCAN.
• Other great SSL repositories: MoCo, SupContrast, SeLa, SwAV and many more here.

License

This software is released under a creative commons license which allows for personal and research use only. You can view a license summary here. Part of the code was based on MoCo. Check it out for more details.

Acknoledgements

This work was supported by Toyota, and was carried out at the TRACE Lab at KU Leuven (Toyota Research on Automated Cars in Europe - Leuven).