Robotic Perception in diverse domains such as low-light scenarios remains a challenge, even upon the employment of new sensing modalities like thermal imaging and specialized night-vision sensors. This is largely due to the high difficulty in obtaining labeled data for certain tasks. In this work, we provide a pathway toward solving two critically important, representative problems corresponding to varying difficulty in obtaining labeled data in this setting, semantic segmentation and object detection. For the more challenging setting, where generating large quantities of new labels can be prohibitively expensive, we propose a novel label-efficient, and effective Domain Adaptation framework that critically accounts for biases learned by the original model in the source domain, and show it on semantic segmentation. Our method outperforms state-of-the-art across a range of visual domains, with improvements of up ~+40% in mIoU in unsupervised, and ~+30% in mIoU in semi-supervised scenarios, in addition to a marked increase in robustness. We also introduce the first ‘intensified’ dataset captured at night time comprising images from an intensifier camera, and a high-sensitivity camera to facilitate low-light robotic operations.

Prof. Jeff Schneider (co-advisor)

Dr. Christoph Mertz (co-advisor)

Zoom Link:  https://cmu.zoom.us/j/96302617425

Meeting ID: 96302617425
Passcode: 622612