Original title: Revisiting Cross-View Completion: Self-Supervised Pre-Training via Reconstruction Error Comparison
Thibaut Loiseau, Guillaume Bourmaud, Vincent Lepetit · September 2026
Computer models can learn about the 3D world by looking at pairs of pictures. They try to fill in missing parts of one picture by checking a second view. However, parts that only appear in one image do not get help from the second view, which weakens training. Researchers created Gekko to solve this issue. Gekko compares two tasks: filling in missing parts using only one picture versus using both pictures. If the second picture helps a lot, the model knows both cameras saw that spot. If it barely helps, only one camera saw it. Gekko predicts this performance difference for every pixel without needing 3D labels. It beat an earlier model called CroCo on several 3D vision tests. It can even train directly on raw videos, skipping the slow data cleanup that older methods needed.
Imagine sorting a deck of cards into two piles: ones that are identical in two different games and ones that are unique to each game. It's hard to tell them apart just by looking at one game at a time. In this paper, the system plays two reconstruction games—one using both views and one using just one—and compares how well it does. By seeing where the errors differ, it figures out which cards are common to both games.
For developers of 3D technologies like self-driving cars or AR glasses, this method makes training simpler and more accurate without needing labeled 3D data. It reduces preprocessing steps, which could speed up development and improve performance in real-world applications.
The abstract doesn't state specific limitations, but it focuses on improvements over CroCo and enables training on raw videos without addressing broader generalization or biases.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Revisiting Cross-View Completion: Self-Supervised Pre-Training via Reconstruction Error Comparison — Thibaut Loiseau, Guillaume Bourmaud, Vincent Lepetit
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.