This repository contains notebooks implementing a weakly supervised approach for detecting animal trespassing in agricultural and residential areas using CCTV images. The method overcomes the lack of bounding box annotations and class imbalance between day and night scenarios through a combination of Class Activation Maps (CAMs), a Discriminative Region Suppression (DRS) module, and adversarial domain debiasing.
- Sri Sethu Madhavan S
- Mohamed Hassan K
- Bhuvaneswari S
All implementation is in
.ipynbformat for better reproducibility and explanation of intermediate outputs.
The detection pipeline addresses the following problems:
- Lack of annotated bounding boxes in real-world surveillance datasets.
- Heavy bias toward daytime images causing poor performance on night images.
- Need for reliable object localization using weak supervision.
Images captured via CCTV are predominantly from daytime, resulting in lighting-based domain imbalance. No bounding box annotations are provided.
We use a VGG-based network with Discriminative Region Suppression (DRS) to expand the attention from highly discriminative parts to the full object extent.
Class Activation Maps are thresholded and converted into bounding boxes through connected component analysis. These serve as pseudo-annotations.
Each pseudo-annotated region is passed through a ResNet-50 classifier. Only high-confidence, semantically correct regions are retained.
A Gradient Reversal Layer (GRL) and domain classifier are introduced into Faster R-CNN to learn domain-invariant features. This ensures that detection performance remains consistent across both day and night images.
KL Divergence is used to monitor prediction divergence across domains. Lower divergence after training confirms effective debiasing.
Clone the repository:
git clone https://github.com/sri-sethu/Debiasing-Object-Detection-Model.git
cd Debiasing-Object-Detection-ModelInstall Dependencies
pip install -r requirements.txt