Action Recogntion

Dataset: HO-3D

Accuracy: 96%

Ho‑3D Action Recognition with Action Transformer

A PyTorch implementation for per‑frame action classification on the Ho‑3D dataset using a lightweight Transformer encoder.

Features

• Parses and annotates Ho‑3D’s object_pose.txt & action_class.txt
• Balances classes via oversampling
• Scales pose features to [0,1]
• Trains an nn.TransformerEncoder‑based classifier
• Reports train/val loss & accuracy per epoch
• Outputs final test accuracy

Requirements

• Python 3.7+
• PyTorch 1.7+
• scikit‑learn
• imbalanced‑learn
• pandas, numpy

Installation

git clone https://github.com/you/ho3d-transformer.git
cd ho3d-transformer
pip install -r requirements.txt