Jetson Nano Dataset Capture Tool Development

I need a small software tool that runs on an NVIDIA Jetson Orin Nano to capture, organize, and export a labeled image/video dataset for training YOLO object detection models.

The camera is an 640×512 sensor connected via MIPI CSI-2 (4 lanes) using a 22-pin ribbon cable to the Jetson Orin Nano CSI connector. The tool should support live preview, capture, dataset management, and YOLO-format export.

Build an application that:
- Interfaces with the camera on Jetson Orin Nano
- Acquire frames from the MIPI CSI-2 camera reliably (GStreamer / V4L2 / libargus or best approach for this sensor).
- Provide live preview (grayscale or false color optional).
- Display FPS + resolution.

Dataset capture:
- Capture single frames and/or short clips (optional).
- Save images in a consistent format (e.g., PNG or TIFF preferred; 16-bit support if available, otherwise 8-bit).
- Support timestamped filenames and session folders.
- Allow adding metadata per session (object name, environment notes, distance, temperature, etc.) saved as JSON/CSV.

Annotation workflow

Option A (preferred): Built-in lightweight annotation UI (draw bounding boxes, assign class).
Option B: Export images to a structure compatible with common label tools (LabelImg / CVAT), then re-import labels.

Must support multiple classes and multiple boxes per image.

Export in standard YOLO format:
- images/train, images/val, labels/train, labels/val
- .txt label files with normalized bbox coordinates
- generate a data.yaml (class names, paths)
- Include train/val split with configurable ratio.

Documentation + setup
Clear instructions for installing dependencies and running on Jetson Orin Nano.
Source code in a Git repo + a short README.
Provide a short demo video or screen recording of the workflow.

Technical Requirements
- Target platform: NVIDIA Jetson Orin Nano (JetPack Linux) - work remotely through ssh
- Camera: 640×512, MIPI CSI-2 4 lanes, 22-pin ribbon

Preferred languages: Python (PyQt / Tkinter / OpenCV).

Must run offline (no cloud requirement).


What I Will Provide:

- Jetson Orin Nano hardware access (remote via ssh).

- Camera details / driver info if needed (device node, gstreamer pipeline, etc.).

- List of target object classes.