Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Prediction heatmap
Two-dimensional heatmap generated by the network indicating the probabilities of object corner presence at each spatial position in the image.
Corner pooling
Specialized pooling operation that explicitly encodes corner information by aggregating horizontal and vertical edge features toward a central point.
Embedding vector
Feature vector associated with each predicted corner, used to match the top-left corner with its corresponding bottom-right corner of the same object instance.
Offset prediction
Additional regression that predicts sub-pixel offsets between quantized corner positions and their actual locations to improve accuracy.
Peak detection
Process of extracting candidate corner locations from heatmaps by identifying local maxima above a confidence threshold.
Corner grouping
Mechanism that associates pairs of corners belonging to the same object using the similarity of their embedding vectors and detection scores.
Feature pyramid pooling
Technique that combines features from different resolutions to improve multi-scale object detection in the CornerNet architecture.
Corner refinement
Post-processing step that finely adjusts the locations of detected corners using local contextual information to improve accuracy.
Top-left corner heatmap
Specific heatmap dedicated to detecting the top-left corners of objects, generated in parallel to that of bottom-right corners.
Bottom-right corner heatmap
Specialized heatmap for predicting the bottom-right corners of objects, working in conjunction with the top-left corner heatmap.
Corner confidence score
Probability value associated with each predicted corner, indicating the model's confidence in the actual presence of a corner at that position.