Semi-Supervised Video Anomaly Detection (Pixel Play 2026)

Hosted by: Vision and Language Group (VLG), IIT Roorkee
Status: Completed (8 January 2026)

📌 Project Overview

This repository contains the official implementation for the Pixel Play 2026 anomaly detection challenge. The objective is to develop a robust system capable of identifying abnormal events in video surveillance footage—such as unauthorized entries, sudden running, or falls—using a semi-supervised learning framework.

The core challenge of Video Anomaly Detection (VAD) is the lack of anomalous training data. Our system defines "normality" based exclusively on a training set of standard activities and flags any significant deviations during inference as potential anomalies.

🚀 Key Features

• Two-Stream Architecture: Simultaneously processes the entire scene (Global) and individual actors (Local) to detect both environmental and behavioral anomalies.
• Foundation Model Backbone: Leverages CLIP (ResNet-101) to extract rich, semantic features that capture high-level concepts better than traditional ImageNet-trained models.
• Human-Centric Analysis: Integrates YOLOv5s to isolate and analyze human poses independently from background noise.
• Memory Efficiency: Implements Greedy Coreset Subsampling to retain only the most diverse 1% of feature vectors, optimizing memory usage without sacrificing accuracy.
• Robust Scoring: Uses a 3-seed ensemble strategy combined with Gaussian Temporal Smoothing to ensure stable and noise-free anomaly detection scores.

🏗️ Technical Architecture

The system employs a unified VideoPatchCore framework that forks into two parallel processing streams:

1. The Global Stream (Context)

• Input: Full $224 \times 224$ video frames.
• Function: Models the background environment.
• Use Case: Detects scene-level irregularities, such as a vehicle entering a pedestrian walkway.

2. The Local Stream (Behavior)

• Input: Dynamic crops of human subjects detected by YOLOv5s.
• Function: Models specific human actions and poses.
• Use Case: Detects behavioral anomalies like fighting, falling, or loitering.

Feature Extraction: Features are extracted from Layer 2 and Layer 3 of the CLIP ResNet-101 backbone. These features are fused and globally pooled into compact embedding vectors for efficient comparison.

📂 Repository Structure

|-- Final_codes
     |-- /training_notebook/         # Main pipeline (final submission - Feature Extraction + Memory Bank creation)
     |-- /Final_submission_output    # Output of training notebook run on kaggle notebooks with GPU P100(T4x2 also works well)
|-- Experiments/                     # Most significant experiments with different models
|-- requirements.txt                 # List of Python dependencies
|-- README.md                        # Project Documentation