When Vision Meets Weather: Building Efficient AI That Sees What Others Miss
About this session
Most computer vision systems deployed in the real world share a quiet vulnerability: they were built and benchmarked under clear conditions. The moment rain falls, fog rolls in, or snow starts blowing sideways, performance degrades in ways that rarely show up in lab evaluations. The conventional response has been to throw more parameters at the problem. More layers, more compute, bigger models. We asked a different question: what if the smarter model is the smaller one? This talk draws from two lines of work pushing the boundary of what efficient vision AI can do under real-world weather. OmniRestore, presented at the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops 2026, achieves state-of-the-art image restoration across rain, fog, snow, and low light with just 2.6M inference-time parameters, outperforming the leading unified model by +1.14 dB PSNR and 41% in perceptual quality, at 56.5% fewer parameters. The second line of work confronts blowing snow, one of the most underrepresented and technically punishing degradation conditions in the field, where the dynamic, density-varying, directional nature of the degradation breaks assumptions baked into virtually every existing restoration pipeline. I will be honest about what broke, why our evaluation metrics lied to us, and what blowing snow taught us about the limits of models that only ever trained on weather someone had already named and benchmarked. You will leave with a sharper instinct for the efficiency-performance tradeoff, a clear sense of when pixel-level metrics are the wrong tool, and a framework for building vision systems ready for the weather conditions nobody planned for.
Speaker
Key takeaways
- Bigger is not better; a 2.6M parameter model can outperform architectures with 28M+ parameters when the design is intentional. Parameter efficiency is a engineering discipline, not a compromise.
- Your evaluation metrics may be lying to you ; PSNR and SSIM systematically underreport perceptual quality failure under composite and blowing snow degradations. Know when to switch to LPIPS and why it matters for real-world deployment.
- The hardest weather condition is the one nobody benchmarked ; blowing snow exposes fundamental assumptions baked into standard restoration pipelines. If your model hasn't been tested on dynamic, density-varying degradations, it isn't production-ready.