Reproduce “A Low Memory Footprint Quantized Neural Network for Depth Completion of Very Sparse Time-of-Flight Depth Maps” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiaowen Jiang, Valerio Cambareri, Gianluca Agresti, Cynthia Ifeyinwa Ugwu, Adriano Simonetto, Fabien Cardinaux, Pietro Zanuttigh
Researchers from EPFL, University of Padova, and Sony developed a compact neural network combined with mixed-precision quantization to accurately complete very sparse Time-of-Flight (ToF) depth maps. This solution drastically reduces the model's memory footprint by up to 14-fold for weights, achieving high-quality depth estimation for mobile and resource-constrained devices.
Key insights
- Mixed-precision quantization-aware training effectively reduces memory footprint for depth completion, outperforming uniform quantization by assigning optimal bit widths per layer.
- Predicting residuals to an initial nearest-neighbors interpolated depth map allows the network to focus on refinement, improving accuracy and efficiency.
- A combined loss function, incorporating both depth L1 error and surface normal similarity, is crucial for producing geometrically consistent and high-quality dense depth maps.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.