Tube Classification Dataset Pipeline

Automated dataset collection, annotation, cleaning, and export pipeline for tube classification using an Intel RealSense D435i and MobileSAM.

Overview

This project provides an end-to-end pipeline for:

• Capture: RealSense RGB + depth acquisition with depth stability checks
• Pre-capture data entry: SQLite-backed class/volume/session metadata
• ROI extraction: Depth-based tube localization
• Segmentation: MobileSAM mask generation
• Cleaning: Blur filtering, duplicate detection, bbox quality checks, background removal
• Export: COCO and YOLO dataset export plus session manifests and stats

Key Features

• RealSense D435i support
• Side and top-down capture workflows
• Verification gate for config, disk, camera, and MobileSAM weights
• Session-based capture with recovery support
• Quality thresholds for blur, mask coverage, IoU, and duplicates
• Guided filtering and depth-mask refinement
• Optional background removal
• Dataset exports in COCO and YOLO formats

---

Quick Start

Prerequisites

• Intel RealSense D435i
• Python 3.10 to 3.12
• Windows 10/11
• USB 3.0 connection
• Enough disk space for your dataset

Installation

git clone https://github.com/Yadav108/Dataset_Pipeline.git
cd tube_classification

python -m venv venv
.\venv\Scripts\Activate.ps1

pip install -r requirements.txt

MobileSAM Weights

Download MobileSAM weights and place them here:

models/MobileSAM/weights/mobile_sam.pt

Run the Pipeline

python capture.py

This will:

1. Run the verification gate
2. Prompt for pre-capture class/volume data
3. Start the capture pipeline
4. Segment and clean captures
5. Save session outputs and exports

To skip the pre-capture metadata step:

python capture.py --skip-pre-capture

----------------------------------------------------------------------------------------------------------------------------------------------------------------------

Usage

Main Capture Flow

python capture.py

Legacy / Interactive Entry Point

python main.py

Quality Analysis

python analyze_image_quality.py

System Checks

python verify_system.py
python verify_hd_capture.py
python verify_mobilesam.py

Dataset Export

python -m src.export.coco_exporter
python -m src.export.yolo_exporter

----------------------------------------------------------------------------------------------------------------------------------------------------------------------

Configuration

Edit config/config.yaml before running.

Current Defaults

camera:
  width: 848
  height: 480
  fps: 30
  color_width: 1920
  color_height: 1080
  color_fps: 30
  depth_min_m: 0.28
  depth_max_m: 0.42

pipeline:
  stability_frames: 4
  blur_threshold: 52.0
  duplicate_hash_threshold: 12
  min_coverage_ratio: 0.48
  sam_iou_threshold: 0.57
  show_preview: false

Tuning Notes

- Increase blur_threshold to reject more blurry images
- Increase min_coverage_ratio for stricter mask quality
- Increase sam_iou_threshold for stricter segmentation acceptance
- Adjust depth_min_m / depth_max_m for your working distance

Guided Filter Defaults

preprocessing:
  guided_filter:
    enabled: true
    radius: 8
    max_processing_time_ms: 500.0

----------------------------------------------------------------------------------------------------------------------------------------------------------------------

Project Structure

tube_classification/
├── main.py
├── capture.py
├── pre_capture.py
├── config/
│   ├── config.yaml
│   └── calibration.yaml
├── src/
│   ├── acquisition/
│   ├── annotation/
│   ├── cleaning/
│   ├── export/
│   ├── orchestrator/
│   ├── preview/
│   ├── processing/
│   └── reconstruction/
├── dataset/
├── data/
├── docs/
│   ├── guides/
│   ├── troubleshooting/
│   └── archive/
├── models/
├── tests/
└── README.md

----------------------------------------------------------------------------------------------------------------------------------------------------------------------

Output Files

Typical session outputs include:

- RGB images
- Depth frames (.npz)
- Segmentation masks
- Annotation JSON
- Session logs
- COCO / YOLO exports
- Quality reports

----------------------------------------------------------------------------------------------------------------------------------------------------------------------

Troubleshooting

Camera Not Detected

- Use a USB
 3.0 port
- Reconnect the RealSense D435i
- Update firmware with RealSense Viewer
- Verify pyrealsense2 is installed correctly

Low Capture Yield

- Increase lighting consistency
- Tune blur and IoU thresholds
- Check camera distance
- Confirm depth range in config.yaml

Poor Segmentation

- Verify MobileSAM weights path
- Adjust ROI and depth range
- Tune sam_iou_threshold
- Check focus and scene background

Memory Issues

- Keep show_preview: false
- Close other camera applications
- Reduce capture load
- Confirm enough virtual memory

----------------------------------------------------------------------------------------------------------------------------------------------------------------------

Performance Targets

Typical session behavior:

- Stable depth gating before capture
- Fast ROI extraction and segmentation
- Quality-based rejection of bad samples
- Session-level logs and recovery metadata

----------------------------------------------------------------------------------------------------------------------------------------------------------------------

Documentation

- README.md — project overview
- QUICK_START.md — quick reference
- docs/README.md — documentation map
- docs/guides/ — setup and optimization guides
- docs/troubleshooting/ — issue fixes

----------------------------------------------------------------------------------------------------------------------------------------------------------------------

Contributing

Contributions are welcome.

Areas that are still open for improvement:

- Multi-camera support
- Better preview performance
- Additional model support
- Cloud export integration
- Web dashboard

----------------------------------------------------------------------------------------------------------------------------------------------------------------------

License

MIT License

----------------------------------------------------------------------------------------------------------------------------------------------------------------------

References

- RealSense SDK: https://github.com/IntelRealSense/librealsense (https://github.com/IntelRealSense/librealsense)
- MobileSAM: https://github.com/ChaoningZhang/MobileSAM (https://github.com/ChaoningZhang/MobileSAM)
- COCO: https://cocodataset.org/ (https://cocodataset.org/)
- YOLO: https://docs.ultralytics.com/ (https://docs.ultralytics.com/)