Revisiting Contrastive Methods for Unsupervised Learning of Visual Representations

TL;DR

This paper investigates the robustness and generality of contrastive self-supervised learning methods across diverse datasets, proposes minor modifications to improve representations, and demonstrates their effectiveness in various tasks without finetuning.

Contribution

The study reveals that contrastive methods are effective across different dataset biases and introduces simple modifications that enhance learned representations for multiple downstream tasks.

Findings

Contrastive methods perform well across various dataset types.

Minor modifications improve representation quality.

Multi-crop training enables spatially structured representations.

Abstract

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. In this paper, we first study how biases in the dataset affect existing methods. Our results show that current contrastive approaches work 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 with minor modifications. 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. The code and models are available at https://github.com/wvangansbeke/Revisiting-Contrastive-SSL.