🐭 DeepRodent

A Robust and Generalizable Vision Framework for Automated Rodent Monitoring in Experimental Biology

Teerapong Panboonyuen (@kaopanboonyuen)

Department of Computer Science, Khon Kaen University · PBYAIL · Chulalongkorn University

✨ Overview

DeepRodent is a unified, detector-agnostic deep learning framework built for precise, generalizable, multi-task rodent monitoring in laboratory environments. It jointly predicts:

📦 Standard object detection
🔄 Oriented Bounding Boxes (OBB) — captures rotation during grooming, rearing, and huddling
🎨 Pixel-level instance segmentation — accurate boundary masks even under occlusion
🧭 Temporal-consistent behavioral embeddings — stable identity tracking across video frames

Rather than bolting these on separately, DeepRodent fuses them into a single-stage, multi-head architecture — so one forward pass gives you everything downstream biology pipelines need.

🎯 Metric	🥇 DeepRodent	Best Baseline (YOLO12-Seg)	📈 Gain
mAP₅₀	96.2	94.4	+1.8
mAP₅₀₋₉₅	84.6	78.2	+6.4
Cross-domain (Lab C)	91.5	87.2	+4.3
FPS (real-time)	154	156	—

🚀 Quick Start

pip install ultralytics

from ultralytics import YOLO

# 🐭 Load DeepRodent
model = YOLO("DeepRodent_WEIGHT.pt")

# 🔍 Run inference
results = model.predict("your_cage_image.jpg", save=True)

# 🎨 Visualize instance masks + OBB
results[0].show()

📥 Download directly

from huggingface_hub import hf_hub_download

weight_path = hf_hub_download(
    repo_id="kaopanboonyuen/DeepRodent",
    filename="DeepRodent_WEIGHT.pt"
)

🧠 Why DeepRodent?

Challenge in the Lab 🔬	How DeepRodent Solves It 💡
Rodents curl, rear, and huddle	Oriented Bounding Boxes capture true rotation
Axis-aligned boxes overlap during interaction	Instance segmentation isolates each animal precisely
Lighting/cage geometry varies across labs	Cross-domain robustness constraint for generalization
Jittery masks across video frames	Temporal consistency loss stabilizes identity tracking
Need for downstream analytics	Direct output → trajectories, heatmaps, behavior states

🏗️ Architecture Snapshot

Input Frame → Shared Backbone → ┬── Detection Head
                                 ├── OBB Head (θ-aware)
                                 ├── Instance Seg Head
                                 └── Temporal Embedding Head
                                          ↓
                        Trajectory Tracking · Behavior States · Occupancy Heatmaps

Detector-agnostic: plug into YOLOv8 → YOLO12 families and gain +2.6 to +3.1 mAP consistently. 🔌

📊 Benchmarks

Evaluated on a private multi-setting laboratory rodent dataset (~30K annotated frames, 60/20/20 split):

🥇 Outperforms RT-DETR by +6.7 mAP₅₀₋₉₅
🥇 Outperforms Mask R-CNN by +13.5 mAP₅₀₋₉₅
⚡ Maintains 154 FPS — real-time, deployment-ready

Full ablation, cross-domain, and SOTA comparison tables are in the paper.

📁 Repository Links

Resource	Link
💻 Code	github.com/kaopanboonyuen/DeepRodent
🌐 Project Page	kaopanboonyuen.github.io/DeepRodent
🤗 Weights (this repo)	`DeepRodent_WEIGHT.pt`
🎮 Interactive Demo	Hugging Face Spaces

🐾 Ethical Use

DeepRodent is an assistive research tool — it does not replace veterinary oversight or trained behavioral experts. Developed following ARRIVE guidelines and the 3Rs principles (Replacement, Reduction, Refinement). Please validate performance before deployment in a new laboratory setting.

📜 Citation

If DeepRodent helps your research, please cite:

@article{panboonyuen2026deeprodent,
  title   = {DeepRodent: A Robust and Generalizable Vision Framework for Automated Rodent Monitoring in Experimental Biology},
  author  = {Panboonyuen, Teerapong},
  year    = {2026}
}

Made with 🧡 for open, reproducible science.

⭐ If this helped your research, consider starring the GitHub repo!

Downloads last month: 15

kaopanboonyuen
/

DeepRodent