This repository is a curated collection of research papers focused on "Computer Vision in the Wild" (CVinW), an emerging field aiming to develop foundation models that can adapt to diverse visual tasks with low cost. It serves researchers and practitioners interested in transferable vision models, offering a structured overview of key concepts, benchmarks, and relevant literature.

How It Works

CVinW is defined by two core principles: broad task transfer scenarios and low task transfer cost. The former involves models that can handle variations in input image distributions and output concept sets, moving beyond traditional closed-set or domain generalization settings. The latter emphasizes efficient adaptation through techniques like few-shot learning and parameter-efficient fine-tuning, enabling models to perform well on new tasks with minimal data or computational resources.

Quick Start & Requirements

This repository is a literature collection, not a software package. No installation or execution is required. It provides links to papers, code, and benchmarks for further exploration.

Highlighted Details

• Comprehensive categorization of papers across various CVinW tasks: Image Classification, Object Detection, Segmentation, Video Classification, Grounded Image Generation, and Large Multimodal Models.
• Detailed sections on papers related to efficient model adaptation, including parameter-efficient methods like prompting and adapters, and other adaptation strategies.
• Extensive coverage of out-of-domain generalization research, including surveys and specific methods for robust model development.
• Links to relevant benchmarks like ELEVATER and challenges such as SGinW, RF100, ICinW, and ODinW.

Maintenance & Community

The collection is actively maintained, with a call for community contributions via issues or pull requests to add missing papers and resources. Links to related reading lists and citation information for key papers are provided.

Licensing & Compatibility

The repository itself is a collection of links and does not have a specific software license. Individual papers and code repositories linked within the collection are subject to their respective licenses.

Limitations & Caveats

As a literature collection, it does not provide executable code or pre-trained models. The breadth of the field means that coverage, while extensive, may not be exhaustive. Users are directed to individual paper links for specific implementation details and requirements.