Erik Strand presents his Master's thesis "Efficient Integration of Lightweight Semantic Segmentation and Depth Estimation for On-Device Video Analytics"

Examiner: Mikael Nilsson

Supervisors: Viktor Larsson, Gustav Hanning

Abstract:

Network cameras observe physical spaces but lack an inherent understanding of the scenes they see. This thesis presents a training-free pipeline for floor and wall instance segmentation from a single uncalibrated RGB image, requiring no camera intrinsics, depth sensors, or model training on labelled floor and wall data. The method combines monocular geometry estimation from MoGe with semantic segmentation from Mask2Former, clustering surface normals by orientation, splitting co-oriented regions by projected depth, and classifying each candidate region via a semantic majority vote.

The pipeline is evaluated on 34 annotated indoor scenes using Intersection over Union (IoU), which measures overlap between predicted and ground-truth regions. The strongest backbone combination tested achieves a floor instance IoU of 0.945 and a wall instance IoU of 0.854 across all 34 scenes. Real scenes captured with Axis network cameras score higher (0.975 floor, 0.900 wall) than the synthetic Hypersim subset (0.938 floor, 0.844 wall). Wall detection is more sensitive to backbone choice than floor detection. The detected instances support downstream applications such as room-boundary estimation, object detector validation, and camera placement verification, all from a single uncalibrated image.