This project investigates the social behavior of adult zebrafish (Danio rerio) using a deep learning-based object detection and tracking algorithm. The study leverages video analysis and region-based metrics to quantify social preferences in controlled laboratory conditions.
- To automatically detect, track, and analyze the movement of zebrafish in a test tank.
- To identify zones of social preference based on spatial and temporal metrics.
- To compute close encounter probabilities using bounding box distances and track IDs.
- Test Tank: Divided into distinct zones (e.g., Region C and Region D).
- Stimulus: Fish in adjacent tanks serve as social stimuli.
- Input: Surveillance video recorded during experiments.
- Output: Annotated video with counts, bounding boxes, and heatmaps.
- Format:
.mp4 - Frame Rate: 5 FPS
- Resolution: e.g., 720p
- Model: Ultralytics YOLOv8
- Framework: Roboflow Inference
- Object detection and fish ID tracking applied across video frames
| Parameter | Details |
|---|---|
| Platform | Google Colab (Free Tier) |
| Language | Python 3.10+ |
| Frameworks Used | Ultralytics YOLOv8, Roboflow SDK |
| Training Location | Roboflow Hosted Training |
| Inference Location | Google Colab + Roboflow Hosted Inference |
| Minimum RAM Required | 4–8 GB RAM (Colab Free Tier works well) |
| GPU Requirements | Optional (for local training); not required with Roboflow |
| Dataset Used | Fish Motions Dataset (annotated via Roboflow) |
| Classes | Fish A1, Fish A2, Fish A3, Fish B |
| Parameter | Value |
|---|---|
| Model Type | Roboflow 3.0 Object Detection (Fast) |
| Pretrained On | COCO Dataset |
| Backbone | CSPDarknet-like |
| Detection Method | One-stage Object Detector (YOLO-style) |
| Model Hosted On | Roboflow Cloud |
| Export Formats | ONNX / TorchScript / CoreML / TensorFlow Lite |
| Input Image Size | 416×416 or 640×640 (auto-scaled) |
| Confidence Threshold | 50% |
| IoU Threshold (NMS) | 50% |
| Parameter | Value |
|---|---|
| Epochs | ~300 |
| Optimizer | SGD (Roboflow default) |
| Learning Rate | 0.01 (decayed) |
| Weight Decay | 0.0005 |
| Batch Size | 16 (approx.) |
| Early Stopping (Patience) | Enabled |
| Metric | Value |
|---|---|
| mAP@50 | 88.5% |
| Precision | 94.7% |
| Recall | 81.0% |
| Class | AP (%) |
|---|---|
| Fish A1 | 100.0 |
| Fish A2 | 61.0 |
| Fish A3 | 94.0 |
| Fish B | 100.0 |
| Average | 88.5% |
- Box Loss: Reduced from
2.8 → 1.5 - Class Loss: Dropped from
4.5 → 1.0 - Objectness Loss: Stabilized around
1.3 - mAP Curve: Converged smoothly after ~100 epochs
- Overfitting: Not observed during training
👉 Click to Watch on Google Drive
👉 Click to Watch on Google Drive
This project demonstrates the successful application of YOLOv8-based object detection for analyzing social preferences in adult zebrafish under controlled laboratory conditions. The use of deep learning, region-based tracking, and Roboflow-powered annotation enabled accurate detection and behavioral inference from video data.
By integrating tools like Ultralytics, OpenCV, and Google Colab, we established a reproducible and scalable workflow suitable for biological research and motion analysis.
Contributions are welcome! Whether it’s improving detection accuracy, adding more behavioral metrics, or enhancing visualization — feel free to fork the repo and submit a pull request.
If you're a researcher working with fish behavior or similar video-based datasets, we’d love to collaborate!
For queries, collaborations, or suggestions:
- 📧 Email: ds6918821@gmail.com
- 🧠 GitHub Issues: Feel free to open an issue for any bug or improvement idea
- Roboflow for dataset management and training
- Ultralytics YOLOv8 for real-time detection
- Google Colab for free training/inference
- Special thanks to our lab team and dataset annotators!
🔍 This work is a part of our ongoing research on automated behavioral analysis of aquatic life.