Skip to content

AGENTS.md

Latest commit

 

History

History
68 lines (50 loc) · 3.59 KB

AGENTS.md

File metadata and controls

68 lines (50 loc) · 3.59 KB

AGENTS.md

Summary

This is the repository reference for matching mismatched camera and NN aspect ratios while preserving as much field of view as possible. Use it when you need the full-FOV resize-mode tutorial rather than just a fixed display-mapping example.

Use This Example When

  • You need to compare crop, stretch, and letterbox strategies.
  • You want a runtime-switchable resize-mode demo.
  • You need a focused tutorial on image-manip resize behavior before inference.

Do Not Use This Example When

  • You need only one fixed resize mode.
  • You need higher-resolution annotation mapping rather than resize-mode switching.
  • You need a minimal detector example without resize experimentation.

Quick Facts

  • Category: tutorials/full-fov-nn
  • Shape: multi-entrypoint+standalone
  • Primary task: run YOLOv6 on full-FOV camera frames with different resize strategies
  • Entrypoints: main.py, cropping.py, letterboxing.py, and stretching.py
  • Standalone path: oakapp.toml packages main.py
  • Frontend: none
  • Runs on: RVC2 peripheral, RVC4 peripheral, and RVC4 standalone packaging
  • Requires: YOLOv6 model assets in depthai_models/
  • Input: live CAM_A
  • Output: Visualizer topics such as Resized, Visualizations, and Resize mode
  • Models: YOLOv6 YAMLs in depthai_models/
  • Visualizer / UI: DepthAI Visualizer via dai.RemoteConnection

Read First

Architecture

  • main.py captures a large 3000x2000 stream and pushes resize-mode changes into ImageManip at runtime through utils/resize_controller.py.
  • The fixed-mode scripts use 812x608 input and hardcode one resize mode each.
  • All variants feed the resized branch into YOLOv6 and publish the NN passthrough frame plus detection overlays.
  • combine_outputs.py exists as an extra helper/example artifact, but main.py does not use it in the current repo state.

Constraints

  • The runtime keybindings in utils/resize_controller.py are a -> LETTERBOX, s -> STRETCH, and d -> CENTER_CROP.
  • oakapp.toml only packages main.py, not the fixed-mode scripts.
  • This tutorial is about resize tradeoffs, so some modes intentionally distort or crop the source frame.

Related Examples

Validation

  • Interactive mode: python3 main.py
  • Fixed crop mode: python3 cropping.py
  • Success looks like: the Visualizer shows resize-mode changes affecting the detection view, and the fixed-mode scripts each stay in their declared strategy
  • Common failure meaning: the README keybinding text was trusted over the current code, or the user expects a generic detector example rather than a resize-technique tutorial