Hi Everyone, we noticed some competitors had had issues with the ‘Hermetic_Small_Vehicle’ feature in the metadata file, so we wanted to provide a small reminder regarding the annotation of small vehicles in the competition dataset.
It is clear that the number of small vehicles in reality (and therefore in the competition imagery) is enormous compared to the other object classes. Since we attempted to avoid an imbalanced dataset, small vehicles were not annotated in some of the frames.
In order to address this issue, we added a feature named ‘Hermetic_Small_Vehicle’ in the metadata file that indicates whether the frame has been hermetically annotated for the small vehicle class. If the value is “True”, then all the small vehicles in the frame are annotated; if “False”, then all the small vehicles in the image are not annotated.
Please note that all the other object classes were hermetically annotated throughout the entire competition dataset.
Anyone have any tips or paper recommendations to read on how to handle the hermetic_small_vehicle = False instances? been researching for a week now on an approach to apply during training on how to handle instances here cars are clearly visible but unlabeled which inherently misleads the learning of the small_vehicle class, but still unclear. any pointer to resources would be helpful :)
Posted by: naivelogic @ March 30, 2023, 8:25 p.m.