Reproduce “DepthFocus: Controllable Depth Estimation for See-Through Scenes” 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
Junhong Min, Jimin Kim, Minwook Kim, Cheol-Hui Min, Youngpil Jeon, Minyong Choi
Researchers at Samsung Electronics developed DepthFocus, a steerable Vision Transformer framework for controllable depth estimation in see-through scenes. This method dynamically modulates feature processing based on a user-intended focus distance, enabling the perception of multiple, layered surfaces behind transparent or reflective objects, and established new state-of-the-art performance on Booster and Middlebury v3 datasets.
Key insights
- Explicitly steering the network's computation with a physical control variable enables active, intent-driven depth perception, allowing the system to query and reconstruct specific depth layers.
- Dynamically modulating features via C-MoE and DCI overcomes representational bottlenecks in multi-layer depth estimation, enabling the network to effectively disentangle overlapping geometries.
- Grounding the control variable to physical depth and using a reference-driven ground truth assignment during training teaches the network to exhibit stable depth estimates across intervals and sharp, deterministic transitions between layers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.