The ECUST Food Dataset is a valuable resource for dietary assessment, yet it lacks the instance-mask annotations needed to train and evaluate instance-segmentation models. In this work, we develop an auditable instance-mask layer encompassing 6,062 polygon masks across 2,978 images, generated by refining source bounding boxes and prompting a frozen Segment Anything Model (SAM). Using this enhanced dataset, we trained single-stage YOLOv8 and YOLO26 segmentation models and compared their performance against two-stage baselines. These YOLO models achieved highly competitive downstream metrics: YOLOv8n-Seg reached a macro volume MAPE of 24.31% and a calorie MAE of 40.29 kcal, while YOLO26n-Seg yielded 25.62% and 41.60 kcal, with both maintaining at least 99.3% evaluation coverage. Crucially, they demonstrated significant efficiency gains, delivering median latencies of 17.8ms (YOLOv8n-Seg) and 22.5ms (YOLO26n-Seg) on a laptop GPU, compared with 442ms for the Faster R-CNN plus SAM baseline. These results validate the proposed mask layer's utility for segmentation training and show that single-stage models reach similar or lower estimation error at a fraction of the latency.  

暂无