GeoBridge: Evolving Transferable Decoder Topologies for Geometric Vision

Rui Wang1 · Haoyu Zhang1 · Ran Cheng2
1Hangzhou Normal University · 2The Hong Kong Polytechnic University

Evolve once. Reuse decoder topologies across geometric vision systems.

Model family

GeoBridge evolves decoder topologies within a fixed multi-scale scaffold. Residual, sequential, and parallel compositions are encoded as functional trees and selected through multi-objective evolution with the gradient sensitivity-alignment proxy (GSAP).

This repository provides the weights for GeoBridgeNet-A1, GeoBridgeNet-A2, and GeoBridgeNet-A3. Each target system reuses the selected topology and primitive configurations, with separately trained weights and prediction interfaces. The bundle covers monocular depth estimation, Gaussian novel view synthesis, and COS3D scene refinement and semantic grounding.

Earlier releases used the name DDN. Checkpoint filenames and architecture keys (A1, A2, A3) retain their existing layout. The code uses the ddn package and some legacy DDN-A1/A2/A3 command-line labels for compatibility.

Download weights

COS3D scenes: Figurines · Teatime · Ramen · Waldo Kitchen.

python -m pip install -U huggingface_hub

# Download the full bundle.
hf download kujimili/GeoBridge-NAS --local-dir ./weights/public

# Download only the NYU GeoBridgeNet-A3 checkpoint.
hf download kujimili/GeoBridge-NAS nyu/a3.pth --local-dir ./weights/public

Evaluation

Use the matching architecture and task checkpoint with the scripts in the code repository. Depth checkpoints include the student encoder. Novel-view synthesis and COS3D use separate environments described in their guides.

Depth evaluation · Novel view synthesis · COS3D guide

The code README documents manuscript results. Follow the task-specific evaluation protocols when comparing results.

All files · Training and search · Component licenses