search

zcablii / SARDet_100K

Offical implementation of MSFA and release of SARDet_100K dataset for Large-Scale Synthetic Aperture Radar (SAR) Object Detection
Other
328 stars 25 forks source link

There may be some annotation errors #3

Closed NeverWangX closed 6 months ago

NeverWangX commented 6 months ago

There are some target annotations with an area of 0 in the annotation file, such as: { “area": 0, "iscrowd": 0, "image_id": 6914, "bbox": [191.0, 1.0, 47.0, 0.0], "category_id": 0, "id": 18603, "ignore": 0, "segmentation": [] } Do you need to clear these annotations with an area of 0

zcablii commented 6 months ago

Thank you for bringing this to our attention! The instances of zero area annotations are a result of the image splitting process. However, these do not impact the training of models because, by default, we filter out instances with areas smaller than 16 pixels in the dataloader config files. This threshold was set specifically to address situations where instances may tear apart during the image splitting process, with zero area instances being a particular case of this scenario. Feel free to safely remove such cases if you prefer, but rest assured, they won't negatively affect the model's training.