Motivation

• OTA considers the label assignment in object detection as an optimal transmission problem. It defines positive/negative training samples for each ground-truth object from a global perspective. 

• The Sinkhorn-Knopp algorithm used in OTA is computationally expensive.

Objectives

• SimOTA applies only the top-k strategy, which reduces additional hyperparameters and maintains the performance. 

Related Works

• The cost between the correct label gi and the predicted pj is calculated using the following formula.

• λ: Balance coefficient　　

• Lcls: Classification error between correct label gi and predicted pj　　

• Lreg: Regression error between correct label gi and predicted pj　　

• For the correct label gi, select k predictions (top-k) in descending order of cost (select from multi positives).

*The value of k varies depending on the correct label.

• Selected predictions were considered positive, and those that were not selected were considered negative. 

Method 2

Experimental Results

Ablations

• RTMDet has modifications on simOTA. (Center Prior Cost)