Published on:
24 September 2023
Primary Category:
Computer Vision and Pattern Recognition
Paper Authors:
Cho-Ying Wu,
Quankai Gao,
Chin-Cheng Hsu,
Te-Lin Wu,
Jing-Wen Chen,
Ulrich Neumann
Collected new dataset with 12 indoor space types to study performance variance
Benchmarked 11 recent methods and revealed bias towards/against certain types
Analyzed 4 training datasets and enumerated biases in their space type coverage
Studied strategies to mitigate imbalance across types
Found generalization to unseen types difficult due to object/scale diversity
Reconsidering indoor depth estimation across different spaces
This paper investigates performance of indoor monocular depth estimation methods across different space types, such as kitchen, classroom, or lounge. It finds severe imbalance between space types for existing methods trained on NYUv2, indicating bias. A new dataset captures 12 space types. Analysis reveals strengths/weaknesses of different training sets. Unseen types are challenging. It emphasizes considering space types for robust deployment.
Monocular depth estimation challenge tests generalization
Realistic room modeling with consumer devices
Demystifying Monocular Depth Estimation with Synthetic Data
Generalized depth inference from RGB and sparse depth
Self-Supervised Monocular Depth Estimation Using Day Images
Unified indoor and outdoor 3D object detection
No comments yet, be the first to start the conversation...
Sign up to comment on this paper