AMB82-MINI: How to dynamically crop detected YOLOv7 bounding boxes

[Clay]

I am working on a project to use a camera module (AMB82-mini) for object detection with YOLOv7. My goal is to capture full-resolution images at 1920x1080 while extracting 640x640 patches for training the YOLOv7 model. The idea is to have high-resolution images NN viewing and use cropped sections (640x640) for neural network training.

I couldn’t find an easy or clear way in the existing AMB82-MINI examples or Arduino SDK to dynamically crop these detected regions from the image.

I tried the method provided in the link below but encountered numerous errors

AMB82 MINI Cut image with face coordinates

Hi @kzezel, You can consider taking one of the examples in open-source Standard SDK as a reference (https://github.com/ambiot/ambpro2_sdk/blob/main/project/realtek_amebapro2_v0_example/src/mmfv2_video_example/mmf2_video_example_vipnn_rtsp_init.c), this is an object detection demo, but this method should work for face detection also. There’s an example on how to do cropping and resizing of your original video. [image] You can set use_roi parameter to 1 and your video will be cropped according…

Is there a better way for doing this?

[Clay]

• Is it possible to manually crop and encode sub-images (regions of interest) to JPG format and store them to SD card using the AMB82-MINI SDK?
• Are there any examples or API calls that support cropping and saving regions as individual JPEGs?
• Can NNObjectDetectionImage.h run inference on image files stored on SD card (e.g., image1.jpg, image2.jpg, etc.)?