I use DL workbench to increase my model performance. I have dataset and 4 key points for each picture on it. I want to find out how much the quality of the model will drop after quantization. To find it out I need my dataset to be suitable to one of formats that DL workbench support. But I didn't find suitable task here:
Supported Dataset Tasks: ImageNet: Classification, Inpainting, Style Transfer Pascal VOC: Object Detection, Semantic Segmentation, Inpainting, Style Transfer COCO: Object Detection, Instance Segmentation, Inpainting, Style Transfer Common Semantic Segmentation: Semantic Segmentation, Inpainting, Style Transfer Common Super-Resolution: Super-Resolution, Inpainting, Style Transfer LFW: Face Recognition VGGFace2: Facial Landmark Detection
But I know, that COCO can be used for key points detection. Is there any way to make COCO or other dataset formats suitable for my task?
Hello Anvar Ganiev,
Thank you for reaching out to us.
We suggest you look at this COCO Dataset on key point detection.
Intel has several pre-trained key points models that might interest you.
Hello Anvar Ganiev,
This thread will no longer be monitored since we have provided a solution. If you need any additional information from Intel, please submit a new question.
You can prepare your dataset in COCO format, which contains key points. More info about keypoints annotation type you can find here: https://www.immersivelimit.com/tutorials/create-coco-annotations-from-scratch, https://cocodataset.org/#keypoints-2018
But for the accuracy measurement for your case (to compare between your and optimized model), you need to edit the accuracy config. For example, you can use accuracy config from OMZ models (from the human poses category, ex: https://github.com/openvinotoolkit/open_model_zoo/blob/master/models/public/higher-hrnet-w32-human-p...). The main parameters are Adapter and Metrics. But each task has its own specific. Can you provide more details about which points are detected in your case?
Thanks for response. I need to detect 4 corner points of car license plates, I have pretty large dataset and coordinates for each image, but not in COCO or other formats. But at the moment I'm working on the more prior for me task, so I'll come here later and let you know if I could deal with it.