CroCo: Self-Supervised Pretraining for 3D Vision Tasks by Cross-View Completion
Philippe Weinzaepfel,
Vincent Leroy,
Thomas Lucas,
Romain Brégier,
Yohann Cabon,
Vaibhav Arora,
Leonid Antsfeld,
Boris Chidlovskii,
Gabriela Csurka,
Jérôme Revaud
Masked Image Modeling (MIM) has recently been established as a potent pre-training paradigm. A pretext task is constructed by masking patches in an input image, and this masked content is then predicted by a neural network using visible patches as sole input. This pre-training leads to state-of-the-art performance when finetuned for high-level semantic tasks, e.g. image classification and object detection. In this paper we instead seek to learn representations that transfer well to a wide variety of 3D vision and lower-level geometric downstream tasks, such as depth prediction or optical flow estimation. Inspired by MIM, we propose an unsupervised representation learning task trained from pairs of images showing the same scene from different viewpoints. More precisely, we propose the pretext task of cross-view completion where the first input image is partially masked, and this masked content has to be reconstructed from the visible content and the second image. In single-view MIM, the masked content often cannot be inferred precisely from the visible portion only, so the model learns to act as a prior influenced by high-level semantics. In contrast, this ambiguity can be resolved with cross-view completion from the second unmasked image, on the condition that the model is able to understand the spatial relationship between the two images. Our experiments show that our pretext task leads to significantly improved performance for monocular 3D vision downstream tasks such as depth estimation. In addition, our model can be directly applied to binocular downstream tasks like optical flow or relative camera pose estimation, for which we obtain competitive results without bells and whistles, i.e., using a generic architecture without any task-specific design.
Cross-view Completion (CroCo, in short) is a self-supervised pretext task consisting of feeding two images of the same scene, one of them being partially masked, to a network. The goal of the pretext task is then for the network to recover the masked pixels. Since the two views have different viewpoints, this is only possible if the network “understands” the 3D structure of the scene, the camera poses and the visual correspondences between the two images.
We present below some reconstruction examples from CroCo on scenes unseen during training. From top to bottom, we show the first image (input), the masked second image (input), the output from CroCo, and the original (ground-truth) second image.
Our CroCo pretext task leads to significantly improved performance for 3D vision downstream tasks, for both monocular and binocular tasks, without bells and whistles, i.e., using a generic architecture without any task-specific design. For instance, we show below some qualitative results for the task of monocular depth estimation.
@inproceedings{croco,
title={{CroCo: Self-Supervised Pre-training for 3D Vision Tasks by Cross-View Completion}},
author={{Weinzaepfel, Philippe and Leroy, Vincent and Lucas, Thomas and Br\'egier, Romain and Cabon, Yohann and Arora, Vaibhav and Antsfeld, Leonid and Chidlovskii, Boris and Csurka, Gabriela and Revaud J\'er\^ome}},
booktitle={{NeurIPS}},
year={2022}
}