How to convert object detection tflite to ambarella bin and anchor bin

[maazullah96]

I converted the YOLOv5.tflite file to the required Ambarella bin format for the CV25 architecture . I think I also need an anchors bin file for running the object detection application on the CV25. Can you suggest how I can generate these two files for object detection? Also, how can I visualize or test the bin file to see if it is able to perform object detection? I tried using AWS SageMaker to convert the file, but I am not sure how to run the file on an object detection app. Is there anything missing in the documentation, or can you provide some guidance on this?

[pataxis]

HI @maazullah96 , thanks for your question on https://github.com/AxisCommunications/acap3-examples/tree/main/object-detection-cv25. We'll get back to you shortly.

[Corallo]

Hello @maazullah96

I am assuming you verified that the yolov5 implementation that you found uses anchors, similarly to how SSD does.

The anchors are strictly dependent on the implementation of the network. The only way to find them is to dig into the repository code where you took the yolov5 model.

If you used the Ultralytics repo, maybe you are looking for these and maybe this can help understand their meaning.

To visualize the output is hard until you don't know how to post process your output, as you would have only the relative position of the boxes to the anchors.

If you want to just check that your model "runs" as it gets loaded in the camera and outputs something, you can just replace the model in the object detection example, changing the size of the input and output tensors accordingly to your model.

[maazullah96]

Thanks, @pataxis and @Corallo for the help. Can you please add or update the documentation of the conversion of the object detection model to the ambarella format as well? because I tried the steps from tensorflow to lard app but my file is outputting only 0 or nan value . I think I am doing something wrong can you guide me about the conversion process I am using the yolov5 medium file and the commands I am using are

source /usr/local/amba-cv-tools-basic.2.5.1.0.2008.ubuntu-18.04/env/cv25.env

/usr/local/amba-cv-tools-basic.2.5.1.0.2008.ubuntu-18.04/cv2/tv2/exe/gen_image_list.py \ -f images -o /home/maaz/Downloads/ambarella_new/img_list.txt -ns -e .jpg -c 0 -d 0,0 -r 256,256 \ -bf /home/maaz/Downloads/ambarella_new/dra_image_bin/ \ -bo /home/maaz/Downloads/ambarella_new/dra_image_bin/draw_bin_list.txt

/usr/local/amba-cv-tools-basic.2.5.1.0.2008.ubuntu-18.04/cv2/tv2/exe/graph_surgery.py tf -p /home/maaz/Downloads/ambarella_new/yolo5-custom-int8.tflite -o /home/maaz/Downloads/ambarella_new/yolo5-custom-int8_iq.pb -isrc 'i:serving_default_input_1|is:1,416,416,3' -t model/tf_upsample_1/resize/ResizeNearestNeighbor,model/tf_upsample/resize/ResizeNearestNeighbor -on "StatefulPartitionedCall"

/usr/local/amba-cv-tools-basic.2.5.1.0.2008.ubuntu-18.04/cv2/tv2/exe/tfparser.py -p "/home/maaz/Downloads/ambarella_new/yolo5-custom-int8_iq.pb" -o yolo5-custom-int8 -of ./yolo5-custom-int8 -d 4 -isrc 'i:serving_default_input_1|is:1,3,416,416'

vas -auto -show-progress /home/maaz/Downloads/ambarella_new/yolo5-custom-int8/yolo5-custom-int8.vas

/usr/local/amba-cv-tools-basic.2.5.1.0.2008.ubuntu-18.04/env/local/cavalry_gen/bin/cavalry_gen -d vas_output/ -f /home/maaz/Downloads/ambarella_new/yolo5-custom-int8.bin

[Corallo]

Unfortunately, we can't provide detailed instructions on specific models. It looks like you are running the gen_image_list using the wrong image size as an argument. You need to provide values to the -r that are equal to the input size of your model. To get further support for the model conversion to cv25 I'd recommend you to contact directly Ambarella.