Cross-Modal Distillation

This project tackles the challenge of estimating depth from thermal imagery where ground-truth data is scarce. Strategy: We leverage the rich geometric feature representations of a pre-trained RGB depth estimation model (the "Teacher") to supervise and train a lightweight Infrared stereo network (the "Student").
Method: By transferring depth knowledge from the data-rich RGB domain, we enable the IR pipeline to learn robust depth estimation without requiring massive annotated thermal datasets.

Spatio-Temporal Infrastructure

To validate this distillation process, I designed and built a custom hardware-software pipeline that bridges the two modalities, ensuring valid data generation for training and evaluation.

• Temporal Alignment (Time Sync): Implemented hardware triggering to achieve microsecond-level synchronization between RGB and Infrared sensors, effectively eliminating motion artifacts in dynamic scenes.
• Spatial Alignment (Calibration): Developed a unified calibration framework to accurately estimate extrinsic transformations between the multimodal sensors, enabling precise reprojection and overlay of visual data.