LOST (Localizing Objects with Self-Supervised Transformers and no Labels) provides a PyTorch implementation for unsupervised object discovery. It targets researchers and engineers seeking to identify and localize objects without labeled data, serving as a crucial step towards fully unsupervised object detection by leveraging novel self-supervised transformer approaches.

How It Works

The core approach utilizes self-supervised transformers, building upon the DINO framework. LOST learns object localization by analyzing visual features derived from self-supervision, bypassing the need for manual annotations. This enables the generation of object maps and bounding boxes, which can then train downstream object detectors without supervision.

Quick Start & Requirements

• Installation: Clone repo, pip install -r requirements.txt. Requires separate DINO framework installation (commit ba9edd1 recommended). Detectron2 (v0.5) needed for object detection extensions.
• Prerequisites: Python 3.7, PyTorch 1.7.1, CUDA 10.2.
• Datasets: PASCAL VOC07/12 and COCO datasets required in datasets folder.
• Links: Paper: arXiv, DINO: GitHub.

Highlighted Details

• Achieves competitive corloc results: VOC07 (61.9), VOC12 (64.0), COCO20k (50.7) with ViT-S/16 and DINO pre-training.
• Extends to unsupervised object detection (Class-Agnostic/Aware) using LOST predictions as pseudo-GT with detectron2.
• Performance varies across ViT/ResNet architectures, DINO/ImageNet pre-training, and patch sizes, detailed in provided tables.

Maintenance & Community

No specific community channels (Discord, Slack) or ongoing maintenance details are provided in the README.

Licensing & Compatibility

Apache 2.0 license. Permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

Strict version dependencies exist for Python (3.7), PyTorch (1.7.1), CUDA (10.2), DINO (specific commit), and detectron2 (v0.5), posing potential compatibility challenges. Setup requires downloading large datasets and complex environment configuration.