Hand object detector for understanding human hands in contact
Top 95.5% on sourcepulse
This repository provides a PyTorch implementation of a Faster R-CNN-based hand object detector, addressing the challenge of understanding human hands in contact with objects from internet-scale data. It is primarily for researchers and developers working on human-computer interaction, robotics, or activity recognition who need to detect and classify hands and their interactions with objects.
How It Works
The detector is built upon the Faster R-CNN architecture, a well-established object detection framework. It leverages a ResNet-101 backbone for feature extraction. The implementation includes specific adaptations for detecting hands and objects, with capabilities to classify hand states (e.g., self-contact, contact with other people or objects) and potentially infer hand side (left/right).
Quick Start & Requirements
handobj_new
), and install PyTorch 1.12.1 with CUDA 11.3. Compile CUDA dependencies by running python setup.py build develop
in the lib
directory. Install Python dependencies via pip install -r requirements.txt
.Highlighted Details
handobj_100K+ego
).handobj_100K
, handobj_100K+ego
) for varying performance characteristics.matching.py
script for post-processing detection results.Maintenance & Community
The project is associated with CVPR 2020 (Oral) and lists Dandan Shan, Jiaqi Geng, Michelle Shu, and David F. Fouhey as contributors. No specific community channels or active maintenance indicators are present in the README.
Licensing & Compatibility
The repository does not explicitly state a license. However, it is based on faster-rcnn.pytorch
(using branch pytorch-1.0
), which may have its own licensing terms. Commercial use compatibility is not specified.
Limitations & Caveats
The project notes occasional false positives with no people present, difficulties with left/right hand classification in egocentric data, and challenges in parsing full states with multiple people. The egocentric models are noted to perform significantly better for egocentric data.
1 year ago
1+ week