[Bug] eval() crashes when an image in the dataset has zero annotations.

[RihaanBH-1810]

When an image with zero annotations is present in the dataset read_annotation() returns []. Passing [] to tv_tensors.BoundingBoxes() creates a tensor with shape (0,) instead of (0, 4). This causes a shape mismatch error downstream during eval().

Pictures supporting the issue :

With coco-2014 dataset :

With coco-2017 dataset :

[dpascualhe]

Hi @RihaanBH-1810 . Thanks for your interest. Can you share the exact data/model used in the evaluation to replicate the bug reported? Thanks!

[RihaanBH-1810]

@dpascualhe , I tried out the val set in both 2014 and 2017 coco dataset
I downloaded them from this site > https://cocodataset.org/#download

As for the model the issue showed up in both yolov5n and yolov8n models
I downloaded the .pt files for the models from the original site and converted them to .torchscript using the ultralytics cli

Yolov5nu.pt from : https://docs.ultralytics.com/models/yolov5/#supported-tasks-and-modes

Yolo8n.pt from : https://docs.ultralytics.com/models/yolov8/#performance-metrics

[RihaanBH-1810]

Also @dpascualhe you might have to rename the coco 2014 images/val2014 to images/val2017 same in the instances json files as well, since currently the tool is hard-coded to open 2017 coco dataset I've opened an issue regarding the same #415