TAO YOLOv3 output


I trained a TAO YOLOv3(MobNetv2 224x224). I get the output as 1x3087x13 (attached is the output from netron).

Can you tell me what is the 13 vector? Like (x,y,w,h…)?

Also in some cases, the class probability score is greater than a 100. Why’s that?

The project is here: mouse vs cup - Dashboard - Edge Impulse

Thanks in advance!


The encoded_detections_out is not the final output. It requires post processing to get the usable boxes, scores and classes.

The last dimension of the model output, in your case 13 is as follows:

(cy, cx, ph, pw, step_y, step_x, pred_y, pred_x, pred_h, pred_w, object, cls...)

To see how we convert this vector into boxes, scores and classes in C++. Export our C++ library for your project and take a look at the fill_result_struct_tao_yolov3..() functions in edge-impulse-sdk/classifier/ei_fill_result_struct.h.

// Raul