PaperReading

ApolloCar3D: A Large 3D Car Instance Understanding Benchmark for Autonomous Driving

January 2020

tl;dr: Detecting 3d pose of a car by detecting keypoints and using Epnp.

Overall impression

ApolloCar3D is a new dataset on ApolloScape. It has ~5k images and ~60k car instances. Around 10 cars per image. They filtered out images with fewer cars from ApolloScape. Compared with Kitti, this dataset has many cars that are distant. It annotates 66 keypoints. (but they also show that the many numbers of keypoints and their semantics are one major source of error in annotation).

Using keypoints significantly improves performance.

Key ideas

Technical details

Notes