I am using YOLO tiny with 300 x 300 preview size. Is there any way to get high quality crop image of detection?
high quality crop Image
Hi Priyam26
Well, no, the detected image will always be <=(300, 300) resolution. You could perhaps try a model with a larger input size (which will increase the latency but will increase the resolution) or try upscaling the cropped image if you really need the detail. Another option would be to retrieve the bounding box for (300,300) preview and then scale it up to the camera resolution and then crop the larger frame (but the preview image must be squeezed, otherwise the bbox coordinates are off).
Hope this helps,
Jaka
Hi Priyam26
If you want to save the detections cropped from high quality frames to .jpg you can take a look at this script.
This is in combination with an object tracker and saving metadata to .csv, but you could modify it for your use case. It is based on class-saver-jpeg, device-nn-sync and stretching-api.