Viet-Anh on Software Logo

What is: Adaptive Training Sample Selection?

SourceBridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection
Year2000
Data SourceCC BY-SA - https://paperswithcode.com

Adaptive Training Sample Selection, or ATSS, is a method to automatically select positive and negative samples according to statistical characteristics of object. It bridges the gap between anchor-based and anchor-free detectors.

For each ground-truth box gg on the image, we first find out its candidate positive samples. As described in Line 33 to 66, on each pyramid level, we select kk anchor boxes whose center are closest to the center of gg based on L2 distance. Supposing there are L\mathcal{L} feature pyramid levels, the ground-truth box gg will have k×Lk\times\mathcal{L} candidate positive samples. After that, we compute the IoU between these candidates and the ground-truth gg as Dg\mathcal{D}_g in Line 77, whose mean and standard deviation are computed as mgm_g and vgv_g in Line 88 and Line 99. With these statistics, the IoU threshold for this ground-truth gg is obtained as tg=mg+vgt_g=m_g+v_g in Line 1010. Finally, we select these candidates whose IoU are greater than or equal to the threshold tgt_g as final positive samples in Line 1111 to 1515.

Notably ATSS also limits the positive samples' center to the ground-truth box as shown in Line 1212. Besides, if an anchor box is assigned to multiple ground-truth boxes, the one with the highest IoU will be selected. The rest are negative samples.