frame

Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!

Sign In

Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!

Tiny-YOLOv2 on NCS outputs wrong BBox values

Hi,

I've trained Tiny-YOLOv2 to detect 3 different classes (blue-,yellow- and orangecones) with darkflow. I extracted a .pb-File from that and tested it out with darkflow, it works.
I converted it to the NCS-format with "mvNCCompile -s 12 -input 416 416 -in input -on output ...". The conversion is successful, but if want to detect the cones in an image it prints out vertical lines.

After that I tried to find out what's happening, I noticed that the width (and sometimes height) value of the BBoxes is near 0, the center seems to be correct.
But this only happens when I use the converted graph on the NCS. The "-ec" flag has no effect on this.

I stored the raw outputs of YOLO on NCS in a file, loaded it in darkflow on the PC and drawed the Bboxes based on this data on the picture, the Bboxes were still lines.
Then I did the opposite, I loaded the data from darkflow on PC in my NCS application to test this algorithm, and now it outputs the correct Bboxes, so the fault must be in the conversion from .pb-file to ncs graph, right?

I took the algorithm for parsing the raw data from YOLO from https://github.com/simo23/tinyYOLOv2 (test.py)
I can upload the .pb and graph files if necessary.

I'd be happy if someone of you has an idea why this is happening and/or how to fix it.

Sign In or Register to comment.