CVPR's NTIRE Challenge Reveals How AI Finally Tackles Raindrop Removal at Night
The second NTIRE raindrop challenge pushed 17 teams to clean adherent water on lenses across day, night, and two focus modes on real photographs.
ProTakeaways
- The NTIRE 2026 challenge benchmarked raindrop removal across day, night, and dual-focus conditions.
- 168 teams registered and 17 submitted valid solutions on 14,139 training and 593 test images.
- AIIA-Lab won with 28.34 PSNR using an MSDT backbone plus scene-level pseudo ground truth refinement.
- Top teams converged on a shared recipe: strong transformer backbone plus test-time pseudo-GT fusion.
- The Raindrop Clarity dataset covers focus on raindrops and focus on background scenes.
- Smallest competitive model was 2.14M parameters, largest reached 953M, showing wide efficiency trade-offs.
Cleaning raindrops off a photo sounds like a solved problem until you try it on a phone that focused on the drops themselves, or on a nighttime street scene where every drop refracts a streetlamp into a glowing halo. That harder version of the task is what the second edition of the NTIRE Day and Night Raindrop Removal challenge at CVPR 2026 tried to benchmark, and the just-released report walks through what worked.
Why raindrop removal breaks existing models
Most public deraining datasets share two blind spots. They consist of images captured by cameras focused on the background, so the raindrops themselves are blurry, and they predominantly contain daytime scenes. A network trained on that data tends to fall over when a droplet is sharp and in focus, or when the only light comes from headlights and neon signs.
The challenge is built on Raindrop Clarity, an ECCV 2024 dataset of real paired shots containing 15,186 high-quality pairs and triplets of raindrop and clean images. The daytime portion holds 5,442 images (3,606 raindrop-focused, 1,836 background-focused) and the nighttime portion holds 9,744 (4,838 raindrop-focused, 4,906 background-focused). Organizers split the data into 14,139 training images, 407 for validation, and 593 for testing.