This repository focuses on Out-of-Distribution (OOD) detection in video action recognition. We propose a novel loss function designed to effectively separate unseen actions from seen ones, enhancing the robustness of action recognition models.
Our model is trained on the UCF101 dataset, achieving:
- 93% accuracy for seen data
- 80%+ accuracy for unseen data
The model utilizes I3D (Inflated 3D Convolutional Network) feature extraction from videos, incorporating both RGB and optical flow inputs.
For a detailed explanation of the methodology, training process, and results, refer to our research paper:
Read the Research Paper
✔️ Custom Loss Function – Ensures effective separation of unseen data from seen classes.
✔️ State-of-the-Art Model – Uses I3D-based feature extraction for superior action recognition.
✔️ Pre-trained Weights Available – Trained on the UCF101 dataset.
✔️ Web Application Interface – Allows users to:
- Provide YouTube & Facebook video links.
- Automatically extract and process videos.
- Perform action recognition and classify actions.
- Identify unseen actions using our OOD detection approach.