Researchers have developed a new framework for semantic segmentation aimed at improving assistive navigation for visually impaired pedestrians. This framework utilizes a novel dataset called SENSATION-DS, featuring chest-height pedestrian-view images with a nine-class taxonomy relevant to navigation. The study evaluated five segmentation architectures, finding that UPerNet-MobileNetV3 achieved the highest mean Intersection over Union, while DeepLabV3Plus-MobileNetV3 demonstrated the lowest error rate for confusing roads with sidewalks and offered practical runtime performance on smartphones. AI
IMPACT This research could lead to more reliable AI-powered navigation tools for visually impaired individuals, improving their independence and safety.
RANK_REASON This is a research paper detailing a new dataset and evaluation of segmentation models for a specific assistive technology application. [lever_c_demoted from research: ic=1 ai=1.0]
- Adarsh Bhandary Panambur
- android
- arXiv
- DeepLabV3Plus-MobileNetV3
- ONNX
- SAM2
- Segment Anything Model 2
- SENSATION-DS
- UPerNet-MobileNetV3
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
