【PaddlePaddle Hackathon 4】RFC of Task 205 Portrait segmentation and Background replacement

[unseenme]

RFC of PaddlePaddle Hackathon 4 Task 205

Solution Name

Portrait segmentation and Background replacement

Description

This notebook shows how to do Portrait segmentation and Background replacement with OpenVINO Runtime and PaddlePaddle model.

Workflow

Models

• portrait_pp_humansegv2_lite

This model is supported by OpenVINO.
The result of benchmark:

...
[Step 11/11] Dumping statistics report
[ INFO ] Count:            55304 iterations
[ INFO ] Duration:         60008.95 ms
[ INFO ] Latency:
[ INFO ]    Median:        4.32 ms
[ INFO ]    Average:       4.29 ms
[ INFO ]    Min:           2.97 ms
[ INFO ]    Max:           35.24 ms
[ INFO ] Throughput:   921.60 FPS

Results Visualizing

Project Timeline

3/20 Submitting RFC
3/31 Creating PR
4/15 Merging PR

My Experience in ML and DL

I'm a PPDE, and I completed three OpenVINO tasks in the first PaddlePaddle Hackathon:
Task 88 openvinotoolkit/openvino#7615
Task 91 openvinotoolkit/openvino#7612
Task 92 openvinotoolkit/openvino#7611

[unseenme]

@zhuo-yoyowz Sorry in your busy schedule. Can this RFC document be passed?

[OpenVINO-dev-contest]

Sorry for late reply and thanks for your submission. From my perspective, I saw a current notebook and your RFC are a little duplicated, could help to explain more on the difference between them.

Meanwhile we still would like to continue collecting more RFC by the deadline of submission, and then select 2 best one from them.

[zhuo-yoyowz]

Yes, I agree with @OpenVINO-dev-contest. New notebooks should be different in the specific use case scenario, DL frameworks, or showing significant improvement over current notebooks, etc. Thanks~

[unseenme]

@OpenVINO-dev-contest @zhuo-yoyowz Thanks for your reviews. Here are the differences I found.

Differences between this solution and notebook 205

	This solution	current notebook 205
DL frameworks	PaddlePaddle	PyTorch
Scenario	dedicated to portrait segmentation	common background removal
How to use the model	load the model directly	convert the model to IR