mks0601 / V2V-PoseNet_RELEASE

Official Torch7 implementation of "V2V-PoseNet: Voxel-to-Voxel Prediction Network for Accurate 3D Hand and Human Pose Estimation from a Single Depth Map", CVPR 2018
https://arxiv.org/abs/1711.07399
MIT License
377 stars 69 forks source link

Will it work to test customized data set using your pre-trained models? #58

Closed happyCodingSusan closed 4 years ago

happyCodingSusan commented 4 years ago

Hi Gyeongsik,

Thank you for your awesome work. Especially thank your for sharing your codes!

The data set Hand2017 is collected with Intel RealSense SR300 camera as indicated by the paper "BigHand2.2M Benchmark:Hand Pose Dataset and State of the Art Analysis".

Assume I collect a new small data set, MY_Dataset, using my own Intel RealSense SR300 camera or Intel Realsense later version SR305 camera. I am sure if I use your pre-trained model for Hand2017 data set will not work on MY_Dataset since camera calibration parameters are not exactly the same. Collecting a new big data set and label it is too time-consuming and it will need special device to label the ground truth of hand poses. I wonder if I can just collect a very small training data set with my camera and label it, and then fine-tune your pre-trained Hand2017 model using my small training set, do you think the fine-tuned model will work on test data set collected with my camera ?

In other words, will transfer learning work for hand pose estimation?

Huge thanks for your kind reply!

mks0601 commented 4 years ago

Hi, I think it gonna work, and actually people use this technique. For example, they pre-train human pose estimation mode on mscoco dataset and fine-tune on posetrack dataset because joint set definition is different.

happyCodingSusan commented 4 years ago

Thank you very much for your quick response!!