Object-centric LeJEPA

2026-07-02 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

Object-centric LeJEPA is a novel self-supervised learning method for image encoders that addresses the data efficiency limitations of image-level approaches. While traditional methods like LeJEPA require extensive datasets, object-centric LeJEPA improves data efficiency by aligning representations at the object level. It overcomes the inherent instability of jointly partitioning and representing objects by utilizing cheap, off-the-shelf SAM proposals for object masks during training. This approach extends LeJEPA's distributional anti-collapse objective to variable-sized sets of objects. An additional instance-separating loss further boosts performance by treating other objects in the same scene as negatives. Evaluated across two model scales and 10-100% of COCO, object-level LeJEPA demonstrates superior performance over image-level LeJEPA on tasks including tracking (DAVIS), classification (ImageNet-1k), segmentation (ADE20k), and re-identification (NAVI).

Key takeaway

For Machine Learning Engineers developing self-supervised image encoders, you should consider object-centric approaches to improve data efficiency. By integrating off-the-shelf object proposals like SAM and an instance-separating loss, you can achieve stronger features with less training data. This method significantly enhances performance on tasks such as tracking, classification, and re-identification, making your models more robust and adaptable across various downstream applications.

Key insights

Object-centric LeJEPA uses SAM proposals to stabilize self-supervised object representation, boosting data efficiency and downstream task performance.

Principles

• Object-level alignment improves data efficiency.
• External proposals can stabilize joint learning.
• Instance separation enhances representation.

Method

Extends LeJEPA's anti-collapse objective to object-centric representations, using SAM proposals for masks. An instance-separating loss treats other scene objects as negatives to boost performance.

In practice

• Apply SAM for object mask generation.
• Integrate instance-separating loss.
• Evaluate on tracking and re-identification.

Topics

• Self-supervised Learning
• Object-centric Representation
• LeJEPA
• SAM Proposals
• Image Encoders
• Computer Vision
• Data Efficiency

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Are Object-Centric Representations Better At Compositional Generalization? — Object-centric representations enhance compositional generalization, especially under data or compute constraints.
• We’re releasing SAM 3.1: a drop-in update to SAM 3 that introduces object multiplexing to significantly improve video processing efficiency without sacrificing accuracy. We’re sharing this update with the community to help make high-performance applications f - x.com — SAM 3.1 improves video processing efficiency via object multiplexing, enabling high-performance AI on smaller hardware.
• Utonia: Toward One Encoder for All Point Clouds — Utonia unifies diverse point cloud domains into a single self-supervised transformer encoder for consistent representation learning.
• Selective Synergistic Learning for Video Object-Centric Learning — SSync selectively distills reliable cues and merges overlapping slots for robust, efficient video object-centric learning.
• SAM 3: Building a unified model architecture for detection and tracking — SAM 3 unifies object detection and tracking using distinct transformers for each task.
• Meta's SAM 3 (Free + Open Source) — SAM 3 enables text-prompted, open-source video object segmentation for easy analysis.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.