This presentation proposes a differentiable ISP adaptation framework for camera migration in
autonomous driving. The central idea is to treat the ISP not as a fixed handcrafted pre-processing
stage, but as a constrained, learnable camera-domain adapter between a new sensor stack and an
already validated perception model. Given a source camera domain used for training and a target
camera domain produced by a new sensor, optics, or ECU tuning, we optimize a differentiable ISP
model that maps target RAW frames into a task-optimal representation while preserving physical
plausibility and deployment constraints. The primary deployment mode keeps the perception model
frozen, so that adaptation occurs only in the ISP layer. A secondary mode allows limited joint
training through lightweight adapters when frozen-model adaptation saturates.
