MVA2023 Small Object Detection Challenge for Spotting Birds: Dataset, Methods, and Results

BandRe: Rethinking Band-Pass Filters for Scale-Wise Object Detection Evaluation

“…This phenomenon is at least partially due to the dataset because it is not happening on COCO. SOD4SB is a dataset for small object detection, which has many small objects and scarce large ones [8]. Thus, a possible cause of low performance in the medium scale range is a combination of performance degradation due to small object scales and a small number of large objects.…”

“…A filter bank with evenly-spaced rectangular or triangular filters could cause some scale ranges with too few objects. For example, datasets for small object detection may have few large objects [8], and the largest or smallest scale range may be narrower than the other scale ranges [15]. In such ranges, scale-wise metrics will be unreliable because they are computed with few specific objects.…”

“…We used models trained on the train2017 split (118,287 images) and evaluated them on the val2017 split (5,000 images). For small object detection, we adopted the SOD4SB dataset [8], a recently released dataset for small object detection for birds. Bird detection is an important application of small object detection for industrial and environmental reasons [17,5].…”