AI 용어집
인공지능 완전 사전
Instance Segmentation
Task that detects and delineates each distinct object in an image, assigning a unique identifier to each instance, without classifying their semantic category.
PQ (Panoptic Quality)
Standard evaluation metric for panoptic segmentation, which weights recognition quality (recog) and segmentation quality (seg) to measure overall performance.
Thing
Category of objects in panoptic segmentation that are countable and have distinct instances, such as cars or people, as opposed to 'stuff'.
Stuff
Category of regions in panoptic segmentation that are amorphous and uncountable, such as sky, road, or grass, as opposed to 'things'.
Pixel Embedding
High-dimensional vector representation for each pixel, used to group pixels belonging to the same instance in instance segmentation approaches.
Detection Head
Module of a convolutional neural network specialized in predicting bounding boxes and object classes, often used upstream of instance segmentation.
Instance Masks
Binary outputs of instance segmentation, where each mask represents the precise shape of a unique object at the pixel level, distinct from other objects.
Unified Panoptic Network
Neural network architecture with multiple heads or a single branch designed to simultaneously generate semantic and instance predictions efficiently.
AP (Average Precision)
Key metric for evaluating instance segmentation performance, measuring the average precision of detections at different Intersection over Union (IoU) confidence thresholds.
Panoptic Segmentation Former (PSFormer)
Transformer-based architecture that treats panoptic segmentation as a mask sequence prediction problem, unifying semantic and instance approaches.
Query-Based Panoptic Segmentation
Paradigm where a set of learned queries interacts with image features to simultaneously predict classes and masks of all objects and regions.
Panoptic Decoder
Final module of a panoptic segmentation network responsible for merging features from semantic and instance branches to produce the final output map.
Panoptic Ground Truth
Image annotation where each pixel is labeled with an instance identifier and semantic class, serving as reference for model training and evaluation.