YOLO26 depth backbone transfers to deraining: beats random init on all 10 test sets

Naive-Explanation940 · reddit · 2026-09-01

A controlled transfer-learning experiment: the author repurposed YOLO26's depth-estimation backbone (CSPDarknet + PAN-FPN) for a different dense-regression task — image deraining — vs. training the same architecture from scratch.

Architecture: backbone/neck unchanged; a new RGBHead restoration decoder reuses the depth decoder's multi-scale fusion, adds a full-resolution reconstruction tail, skip connections from stride-2/4 layers, NAFNet-style residual output, and LayerNorm in the head (backbone keeps BatchNorm for TensorRT export and zoo compatibility). Loading the depth checkpoint matches 468/468 backbone+neck tensors — only the head is random-init, giving a clean controlled comparison.

Result: identical architecture and recipe, fixed 100 epochs at nano scale — depth init averaged 27.94 PSNR / 0.813 SSIM vs. 27.45 / 0.807 random, winning all 10 test sets. The +0.48 dB gap appeared by epoch 20 and didn't close, ruling out a convergence-speed artifact. The author notes this doesn't establish why — geometry from depth supervision or simply a strong checkpoint — and released nano (5.25M) and small (12.13M) models plus per-dataset deltas.

Original post →

More from Research

Research channel →