Objectron: A Large Scale Dataset of Object-Centric Videos in the wild with pose annotations
3DObjectDetection, 이미지에서 depth 추론 아이디어를 얻기 위한 논문 읽기 진행
Introduction
3D object understanding -> object-centric video dataset -> matching the data format for many down-stream vision task
Object-centric approach, Objectron dataset
: 물체 중심의 비디오, which capture different views of the same objects from different angles + 3D bounding box + accompanied by AR sessions metadata
Data Collection and Annotation
Object Categories & Data Collection
Goal: capture objects in their common environment and in relative context
Objectron: A Large Scale Dataset of Object-Centric Videos in the wild with pose annotations
3DObjectDetection, 이미지에서 depth 추론 아이디어를 얻기 위한 논문 읽기 진행
Introduction
3D object understanding -> object-centric video dataset -> matching the data format for many down-stream vision task
Data Collection and Annotation
Object Categories & Data Collection
Data Annotation
: 수작업과 AR을 통한 데이터 annotation
3D IoU(Intersection over union)
다음 ipynb 코드에서 기준을 볼 수 있다.
간단히 말해서 Suther-Hodgman algorithm으로 polygon을 겹치고, convex hull 알고리즘으로 intersection을 계산한다.
컵이나 병과 같은 대칭 개체는 대칭축을 잡고 회전개체를 계산하여 구할 수 있다.