PREGAN: Pose Randomization and Estimation for Weakly Paired Image Style Translation

Utilizing the trained model under different conditions without data annotation is attractive for robot applications. Towards this goal, one class of methods to translate image style from another environment on which models are trained. In paper, we propose a weakly-paired setting translation, where content in two images aligned with errors poses. These could be acquired by sensors that share an overlapping region, e.g. LiDAR or stereo cameras, sunny days foggy nights. We consider more practical with: (i) easier labeling than paired data; (ii) better interpretability and detail retrieval unpaired data. To across such images, PREGAN train translator intentionally transforming random pose, estimate given pose differentiable non-trainable estimator style, estimated result is. Such adversarial training enforces network learn avoiding being entangled other variations. Finally, validated both simulated real-world collected show effectiveness. Results down-stream tasks, classification, road segmentation, object detection, feature matching its potential real https://github.com/wrld/PRoGAN