Re: tensorflow uint8 object detection model

MyungJoo Ham

On Thu, Aug 12, 2021 at 01:41 PM, shortcipher wrote:

Thanks for the quick response!  I tried that modification adding the threshold and permuting the output tensor indices and I'm still getting the same error.  (I also tried all other permutations of the indices 0:1:2:3,50 / 3:2:1:0,50 / 3:1:2:0 / etc) Did you make any other changes to get it to work well?

Here is the gst pipeline description that I've tested:

gst-launch-1.0 v4l2src name=cam_src ! videoconvert ! videoscale ! video/x-raw,width=640,height=480,format=RGB,framerate=30/1 ! tee name=t t. ! queue leaky=2 max-size-buffers=2 ! videoscale ! video/x-raw,width=320,height=320,format=RGB ! tensor_converter ! tensor_transform mode=arithmetic option=typecast:uint8,add:0,div:1 ! tensor_filter framework=tensorflow2-lite model=model.tflite ! tensor_decoder mode=bounding_boxes option1=mobilenet-ssd-postprocess option2=coco-labels-paper.txt option3=0:1:2:3,50 option4=640:480 option5=320:320 ! compositor name=mix sink_0::zorder=2 sink_1::zorder=1 ! videoconvert ! ximagesink t. ! queue leaky=2 max-size-buffers=10 ! videoconvert ! mix.

It appears that the video format (RGB) from t might be not compatible with compositor:

Join { to automatically receive all group messages.