RRC intern

video object detection on BDD100K

This project focuses on detecting objects in driving videos under challenging weather conditions using temporal aggregation techniques.

Predictions on blurred real-world data from BDD100k dataset

Methods & Benchmarks

  • Fine-tuning Faster R-CNN:
    • The Faster R-CNN model was fine-tuned on the BDD100K dataset to improve detection in adverse weather and lighting conditions.
    • Used Detectron2 framework. LINK
  • Temporal Keyframe Matching:
    • Windowed Hungarian Algorithm (No-Learning) for object tracking.
    • LightGlue (No-Learning) for feature matching.
  • Faster R-CNN Benchmarking:
    • Benchmarking was conducted to compare performance against standard models and evaluate improvements.

Goal

The primary goal of this project is to improve object detection accuracy in real-world driving scenarios without adding complex learning.