Yolov3 from public models gives wrong bounding boxes

[Ukhupacha]

Hello, 

I used to follow these instructions to get yolov3.pb and then IR, it worked nicely.

However, I noticed you uploaded YoloV3 and I thank you that, I wonder if you need some extra parameters to convert it to IR.with the model you have uploaded.

I used the optimizer with this command 

python3 mo_tf.py \
	--input_model yolov3.pb \
	--transformations_config yolo_v3_new.json \
	--input_shape [1,416,416,3] \
	--output_dir /yolo/FP32 \
	--model_name yolo_v3 \
	--data_type FP32

But now, I get these weird bounding boxes, see yolo.PNG
I tried SSD and it work fine, see ssd.PNG

Cordially,

Ukhu

[JAIVIN_J_Intel]

Hi Ukhu,

Thanks for reaching out.

I have executed the YoloV3 model on object_detection_demo_yolov3_async python demo, after converting to IR. But the bounding boxes were not present as desired.

We will take a look into this. Meanwhile, please use the model you have obtained by following instructions for Converting YOLO* Models to the Intermediate Representation (IR)

Regards,

Jaivin

[Ukhupacha]

Thanks Jaivin,

Please keep me updated if the yolov3 model you have uploader is corrected.

Ukhu

[Vladimir_Dudnik · Accepted solution]

@Ukhupacha,

I would not recommend you to take models directly from download01.org. Doing so, you may get some pre-released content (like in your case), or some outdated staff, you also will have to guess on model conversion parameters and other details. The model you mention is planned to be included into OpenVINO 2020.2 release, which is coming soon. You'd better access Open Model Zoo models through OMZ downloader and converter scripts. You need use tools from OMZ master branch (or use OMZ from OpenVINO package) if you want to access the latest public release content and you may use OMZ tools/models from develop branch (on your risk) to access current work-in-progress content.

Please look at README for more details on how to use OMZ tools. These scripts will help you to download and then convert model to IR and we do validate this for every OpenVINO release.

Regards,
  Vladimir