Intel® Distribution of OpenVINO™ Toolkit
Community assistance about the Intel® Distribution of OpenVINO™ toolkit, OpenCV, and all aspects of computer vision-related on Intel® platforms.

offset in detection from ssd

pkhan10
New Contributor I
1,007 Views

Hello 
I used mobilenet_ssd_v2 from tensorflow model zoo,
Its working fine for nearby object but for distant object there's offset in output
As you can check in following images
I used openvino ssd async example ....for getting following output

Screenshot from 2019-09-19 11-48-02.png

I used following commands to get model converted...

python mo_tf.py --input_model ../model_downloader/object_detection/common/ssd_mobilenet_v2_coco/tf/ssd_mobilenet_v2_coco_2018_03_29/frozen_inference_graph.pb -o ../model_files/ssd_v2/ --tensorflow_use_custom_operations_config extensions/front/tf/ssd_v2_support.json --tensorflow_object_detection_api_pipeline_config  ../model_downloader/object_detection/common/ssd_mobilenet_v2_coco/tf/ssd_mobilenet_v2_coco_2018_03_29/pipeline.config --reverse_input_channels

0 Kudos
10 Replies
Shubha_R_Intel
Employee
1,007 Views

Dear khandelwal, prateek

Were your model optimizer args like this ? Take a look at C:\Program Files (x86)\IntelSWTools\openvino_2019.2.275\deployment_tools\tools\model_downloader\list_topologies.yml and search for "ssd_mobilenet_v2" Looking at your MO command above it looks different from below. For instance you're not doing - --output=detection_classes,detection_scores,detection_boxes,num_detections

model_optimizer_args:
      - --framework=tf
      - --data_type=FP32
      - --reverse_input_channels
      - --input_shape=[1,300,300,3]
      - --input=image_tensor
      - --tensorflow_use_custom_operations_config=$mo_dir/extensions/front/tf/ssd_v2_support.json
      - --tensorflow_object_detection_api_pipeline_config=$dl_dir/ssd_mobilenet_v2_coco_2018_03_29/pipeline.config
      - --output=detection_classes,detection_scores,detection_boxes,num_detections
      - --input_model=$dl_dir/ssd_mobilenet_v2_coco_2018_03_29/frozen_inference_graph.pb

Hope it helps. 

Thanks,

Shubha

0 Kudos
pkhan10
New Contributor I
1,007 Views

hello shubha..
I tried the command you share to convert model...
but still no change...
pls help..

python mo.py --framework=tf --data_type=FP32 --reverse_input_channels --input_shape=[1,300,300,3] --input=image_tensor --tensorflow_use_custom_operations_config=extensions/front/tf/ssd_v2_support.json --tensorflow_object_detection_api_pipeline_config=../model_files/ssd_v2/ssd_mobilenet_v2_coco_2018_03_29/pipeline.config --output=detection_classes,detection_scores,detection_boxes,num_detections --input_model=../model_files/ssd_v2/ssd_mobilenet_v2_coco_2018_03_29/frozen_inference_graph.pb -o ../model_files/ssd_v2/
 

 

Screenshot from 2019-09-20 15-32-16.png

0 Kudos
Shubha_R_Intel
Employee
1,007 Views

Dear khandelwal, prateek,

Are you using OpenVino's SSD Demo to run inference ?  We also have a Python version of SSD Async . If you are writing your own code from scratch please compare your code to OpenVino samples.

Thanks,

Shubha

0 Kudos
pkhan10
New Contributor I
1,007 Views

Hello shubha..
I used python example only..
The problems is with original model itself..
there's offset for far away objects...
 

0 Kudos
Shubha_R_Intel
Employee
1,007 Views

Dear khandelwal, prateek,

Yes it seems so. I think this may be a bug. Can you attach your photo to this ticket ? I will attempt to reproduce and get back to you on this forum. 

Thanks !

Shubha

0 Kudos
pkhan10
New Contributor I
1,007 Views

hey shubha
what do you mean by photo?
you mean personal pic?

0 Kudos
Shubha_R_Intel
Employee
1,007 Views

Dear khandelwal, prateek

Yes, I mean that very photo you used to find the problem - 

I have PM'd you so that you can share it privately.

Thanks,

Shubha

0 Kudos
Shubha_R_Intel
Employee
1,007 Views

Dear prateek,

I have PM'd you. But I need the actual *.jpg photo you used in this post (without the bounding boxes of course). Also I assume that you don't see these offsets using the mobilenet_ssd_v2 tensorflow model when you use Tensorflow  for inference ? If you're sure that Tensorflow is accurate then this could be an OpenVino bug.

Thanks,

Shubha

0 Kudos
pkhan10
New Contributor I
1,007 Views

Hello shubha..
I said the same, in my last comment
the problem is there with original model itself...
i tested tensorflow model on nvidia gpu machine..
got similar offset..
offcourse..i didn;t compare result exactly..but ...there was visible offset in ssd mobilenet out of box itself

0 Kudos
Shubha_R_Intel
Employee
1,007 Views

Dear khandelwal, prateek,

OK I missed it. Then in this case it's not an OpenVino issue, since you're observing the same offset on an Nvidia GPU. It's a Tensorfow Model issue. Glad we sorted this out ! Unfortunately this forum is to address OpenVino questions, namely Model Optimizer and Inference Engine. Kindly consult the Tensorflow forums for help with Tensorflow specific issues.

Thanks !

Shubha

 

0 Kudos
Reply