README.md

metadata

license: apache-2.0
language:
  - en
tags:
  - computer-vision
  - animal-pose-estimation
  - multi-object-tracking
  - occlusion-handling
  - image-segmentation
  - inpainting
  - deep-learning
  - video-analysis
arxiv: 2512.07712

UnCageNet

UnCageNet is a computer vision framework for robust animal tracking and pose estimation in caged environments, where occlusions caused by cage bars significantly degrade the performance of existing methods.

This repository provides the official implementation of the paper:

UnCageNet: Tracking and Pose Estimation of Caged Animal
Sayak Dutta, Harish Katti, Shashikant Verma, Shanmuganathan Raman
arXiv: https://arxiv.org/abs/2512.07712

🔗 Code: https://github.com/itz-sayak/UnCageNet

---

🔍 Method Overview

UnCageNet introduces a three-stage preprocessing pipeline that improves downstream tracking and pose estimation under structured occlusions:

1. Cage Segmentation

   • Gabor-enhanced ResNet-UNet
   • Orientation-aware filters (72 directional kernels)
   • Accurate detection of cage bar structures

2. Cage Inpainting

   • Content-aware reconstruction using CRFill
   • Removes structured occlusions while preserving animal appearance

3. Downstream Evaluation

   • Standard pose estimation and tracking models (e.g., STEP, ViTPose)
   • Applied on “uncaged” frames for fair performance comparison

This pipeline enables performance comparable to uncaged environments, despite heavy occlusions.

---

📊 Experimental Highlights

• Significant improvement in:
  • Keypoint detection accuracy
  • Trajectory consistency
• Robust performance across:
  • Severe occlusion patterns
  • Long video sequences
• Plug-and-play compatibility with existing tracking and pose models

(Refer to the paper for full quantitative results.)

---

💡 Intended Use

UnCageNet is intended for:

• Animal behavior analysis
• Zoological and veterinary monitoring
• Laboratory animal studies
• Long-term tracking in constrained environments

---

⚠️ Limitations

• Assumes structured occlusions (e.g., cage bars)
• Performance may degrade for:
  • Highly deformable or unstructured occluders
  • Extremely low-resolution video
• Not trained for arbitrary object categories beyond animals

---

📄 Citation

If you use this work, please cite:

@article{dutta2025uncagenet,
  title   = {UnCageNet: Tracking and Pose Estimation of Caged Animal},
  author  = {Dutta, Sayak and Katti, Harish and Verma, Shashikant and Raman, Shanmuganathan},
  journal = {arXiv preprint arXiv:2512.07712},
  year    = {2025}
}