Anupam Rawat CV
← Hackathons

Hackathon · 2024

Qualcomm VisionX

Removing reflections from photos taken through glass, with a model that holds up under adversarial inputs.

Code ↗

Reflection-removal architecture: multi-scale transformer blocks with cross-scale attention, fusion and an adversarial discriminator

Problem

Detect and remove reflections from a single photo taken through a window or glass, while keeping the scene behind it intact. Entries were measured with PSNR and SSIM on clean and adversarially attacked images.

What we built

  • A model built on RobustSIRR (CVPR 2023): transformer blocks with dynamic convolutions, a cross-scale module and multi-scale fusion.
  • Training with perceptual, contextual, gradient and adversarial losses.
  • Training data from VOC2012 and real Berkeley photos; tests on Real20, SIR² and Nature20.
  • An evaluation harness that attacks the model on purpose.

Design choices

  • Training on a mix of clean and attacked images, so the model keeps working when its input is perturbed.

Result

Rank 4 in the challenge.