We train a single model for all objects on each dataset, and based on an architecture of object detection and pose regression.

Object detection: extended FCOS

Pose regression: extended PFA-Pose

Data: PBR + Real (if available)

RGBD track: the same models used in RGB Track, RANSAC-Kabsch for depth utilizing.

The Main differences from FCOS:

1. We use stronger augmentations following the best practice in 6D pose estimation

2. We utilize some mask information for a better sampling of positive signals during training

The main differences from the original PFA-Pose paper:

• It does not use a detection component, we embed a detection component into it to facilitate the pose regression.

• It uses exemplars rendered offline for training, which is resource-friendly and efficient during training. For this competition, we replace it with online rendering to achieve better accuracy.

• It uses only flow from the rendered image to the input. We further use a backward flow from the input to the rendered image for a consistent check to remove more outliers.

List of contributors:

Yang Hai, Rui Song, Zhiqiang Liu, Jiaojiao Li (Xidian University)

Mathieu Salzmann, Pascal Fua (EPFL)

Yinlin Hu (Magic Leap)