Input: $k^2\times \left(C+1\right)$ Position-sensitive score maps Output: Classification confidence scores Steps: 1. Position-sensitive continuous score maps generation: o For each ROI, generate $k^2\times \left(C+1\right)$ position-sensitive score maps. These maps represent the likelihood of an object being present in specific positions within the ROI. 2. RoIs application to score maps: oApply each ROI to the corresponding position-sensitive score maps to extract features. This involves overlaying the ROI on the score maps to gather position-specific responses. 3. PS-Pr-RoI pooling: oPerform PS-Pr-RoI pooling to aggregate the features from the score maps. This involves dividing each ROI into k × k bins and computing the average feature value for each bin without quantizing the coordinates. 4. Voting for classification confidence: oAggregate the pooled features to generate a C+1-dimensional vector for each ROI. This step involves averaging the scores from the k × k bins for each class. 5. Softmax computation for classification confidence: oApply the softmax function to the C+1-dimensional vector to obtain classification confidence scores for each ROI. This converts the raw scores into probabilities that sum to 1. 6. Output classification confidence: o Output the classification confidence scores for each ROI. These scores indicate the likelihood of each ROI belonging to the respective classes.