kotaharsha25/yolov8-image-and-video-processing-application

Purpose and Entry Point

NexVision AI is a Flask-based computer-vision application for YOLOv8 object detection, instance segmentation, and pose estimation across uploaded media and live video sources. The runtime entry point is app.py, launched with python app.py and serving the browser UI on port 7860.

Application Shell and UI Routing

app.py is the backend composition point: it hosts the Flask application, exposes page/API behavior, and connects input sources to inference streaming. The UI is organized by source-specific templates:

• templates/index.html

• templates/local.html

• templates/yt.html

• templates/rtsp.html

• templates/cams.html

The frontend uses browser-side JavaScript/AJAX behavior described in the README to switch sources and models without full-page reloads. Shared visual styling resides in static/css/theme.css.

Input Source Boundary

The system accepts four materially different acquisition modes: local image/video uploads, YouTube streams, RTSP/IP-camera streams, and local webcam feeds. These converge in the backend into a common frame-processing/inference path, while source-specific UI pages collect the relevant URL, device, or uploaded-media input.

Bundled sample media is held under:

• data/raw/temp_video.mp4

• data/raw/video_1.mp4

YOLOv8 Inference Orchestration

The core ML boundary is Python/OpenCV frame handling feeding Ultralytics YOLOv8 models backed by PyTorch. The application supports dynamically choosing detection, segmentation, or pose-estimation models from the UI. Model loading is intentionally lazy according to the README, reducing memory pressure by avoiding eager loading of all model variants—important for constrained hosted deployment.

Primary implementation location:

• app.py

Streaming and Result Delivery

For video-capable sources, control flow is: browser starts a source → Flask acquires frames → OpenCV/YOLO processes frames → annotated output is streamed back for native browser playback. Start/stop behavior is intended to be asynchronous from the user experience, with frontend AJAX separating UI state from inference execution.

This creates two important boundaries: source acquisition may fail independently (camera, RTSP, YouTube), and model inference is the compute-heavy stage that must remain isolated from template rendering and request navigation.

Data and Model Metadata

YOLO dataset/model class metadata is represented by YAML files under data/, including:

• data/coco.yaml

• data/coco128.yaml

• data/coco128-seg.yaml

These files provide detection/segmentation dataset definitions rather than being application UI data stores. There is no database or persistent business-data layer evident in the repository tree.

Runtime, Packaging, and Deployment

The material runtime is Python with Flask, PyTorch, OpenCV, and Ultralytics YOLOv8; dependency installation is defined by requirements.txt. Container deployment is supported by Dockerfile. The README identifies Render as the hosted target and specifically motivates lazy model orchestration to stay within Render memory limits. Utility-level shared backend behavior is separated into:

• utils/general.py