.png){kind=icon}
Return
TCKD: Task-Oriented Contrastive Knowledge Distillation for Object Detection
.png){kind=icon}
PRE
AI
DOI：10.1109/tmm.2026.3673559.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Most existing prediction-based distillation methods for object detection suffer from misalignment between distillation losses and optimization objectives, which hinders the effective transfer of critical knowledge and severely limits the student’s improvement. To address this challenge, we propose a novel task-oriented knowledge distillation (TCKD) method that transfers both classification and regression knowledge through two efficient distillation losses: class-wise contrastive distillation loss (CCDL) for classification and point-wise contrastive distillation loss (PCDL) for regression. Specifically, CCDL constructs positive and negative samples based on class predictions and establishes intra-class consistency and inter-class discrepancy between samples. This guides the model to effectively focus on regions where the objects align with their corresponding classes. PCDL generates positive and negative samples by leveraging sample points at different spatial positions and captures finer-grained localization differences through similarity computation. Moreover, the proposed foreground-focused strategy used for PCDL effectively reduces background noise and alleviates the distribution imbalance between the foreground and background. Extensive experiments on the MS COCO2017, visDrone2019, and TinyPerson benchmarks demonstrate that TCKD not only achieves superior performance on large-scale datasets and challenging small object datasets but is also adaptable to various detectors.
Keywords:
Contrastive learning
knowledge distillation
object detection
prediction imitation
Journal
IF:
9.7
Papers:
4.7K
Citations:
2.4W
Organization
Cited Papers
.png){kind=icon}
ShareSaveShareSave
ShareSaveShareSaveShareSave
View more