Results are different between tensorflow SSD/Faster-RCNN pb and openvino

[lin__bofan]

Hi, 

I trained the Quantized trained SSD with Mobilenet v2 on MSCOCO Dataset and Faster R-CNN with Resnet-101 (v1), configured for Pascal VOC Dataset. And converted to Openvino(2020.2) model succeeded only by ssd_support_api_v1.15.json and faster_rcnn_support_api_v1.14.json respectively. However, the results between pb and ov model are quite different (ov is wrong). The data input to models is equally. I am really confused about what could cause this problem. I think there must be some problems when converting the model. Could anyone give some instructions? Thanks a lot.

bofan

[lin__bofan]

does anyone saw this topic?

[SuryaPrata_C_Intel]

Hi bofan,

The issue seems to be with the json, as the json files are provided for the pretrained models of OpenVINO.

Please edit the json as per your models if you are not using one of the pretrained models provided with OpenVINO.

Best regards,

Surya