liudongdong1 收录于视觉AI

2022-10-21 约 390 字预计阅读 1 分钟 - 次阅读

https://cdn.stocksnap.io/img-thumbs/280h/ZVCXJVXL75.jpg

level: 2020, CCF_A CVPR author: Yang You, Shanghai Jiao Tong University date: 2020 keyword:

3D keypoint

You, Y., Lou, Y., Li, C., Cheng, Z., Li, L., Ma, L., … & Wang, W. (2020). KeypointNet: A Large-scale 3D Keypoint Dataset Aggregated from Numerous Human Annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 13647-13656).

Paper: KeypointNet

Summary

present keypointnet, the first large-scale and diverse 3D keypoint dataset that contains 103450 keypoints and 8234 3D models from 16 object categrories;
propose a novel methods to aggregate these keypoints automatically through minimization of a fidelity loss;
propose two large-scale keypoint prediction tasks: keypoint saliency estimation, and keypoint correspondence estimation; and experiment including point cloud, graph, voxel and local geometry based keypoint detection.
In order to generate ground-truth keypoints from raw human annotations where identification of their modes are non-trivial

Research Objective

Application Area: object matching, object tracking, shape retrieval, registration, pose estimation, matching, segmentation; which is invariant to rotations, scales and other transformations;
Purpose:

Proble Statement

few 3D datasets focusing on the keypoint representation of an object;
different people may annotate different keypoints which need to identify the consensus and patterns;
predefined distance threshold fail to identify closely spaced keypoints;

previous work:

Detection of keypoints:
- traditional methods: 3D Harris, HKS, Salient Points, Mesh Saliency etc extract geometric features as local descriptors, but only consider the local geometric information without semantic knowledge.
- DNN methods: SyncSpecCNN don’t handle rotations well.
Keypoint Datasets:
- Keypoints for human skeletons: MPII human pose dataset, MSCOCO keypoint challenge, PoseTrack;
- Animals: PUB provides 15 part locations on 11788 images from 200 bird categories
- 3D objects:

Methods

Problem Formulation:

given a valid annotation from c-th person, the keypoint set is:

system overview:

【Dataset Visualization】

【Keypoint Saliency】 assume that the each annotation is allowed to be erroneous within small region. $\Phi$ is the Gaussian kernel, and $Z$ is the normalization function;