Reproduce “Characterizing an On-Device Depth-Aware Multimodal Workload in Assistive Navigation for Blind and Low-Vision Users” 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
Aydin Ayanzadeh, Tim Oates
The paper measures an on-device depth-aware navigation pipeline phase by phase. In fixed-output tests, decoding accounts for 43.7–74.5% of latency, while detection and depth account for 3.1–5.9%. Removing structured scene information can also lengthen answers: one configuration produced 29 tokens and 1,036 ms, versus 17 tokens and 695 ms. These laptop-GPU measurements characterize workloads, not navigation safety.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.