HyCo-DeB: Hyperbolic and Contrastive Embedding using Decoupling with Baby Learning for Few-Shot Object Detection
Authors
Vu, Anh Tuan Nhu ; Jankowski, Alexander Pugholm ; Nielsen, Tobias Kastbjerg Hauge
Term
4. term
Education
Publication year
2022
Abstract
At opdage nye objekter ud fra blot få annoterede eksempler – kaldet few-shot objekt-detektion (FSOD) – er eftertragtet, men stadig svært. En central nøgle er, hvordan billedernes visuelle træk (features) repræsenteres som indlejringer (embeddings). Mange systemer bruger et euklidisk rum, selv om et hyperbolsk rum kan gengive hierarkisk information mere naturligt. En anden måde at forbedre indlejringerne på er kontrastiv læring, som trækker eksempler fra samme klasse tættere sammen og skubber forskellige klasser længere fra hinanden. Vi præsenterer HyCo-DeB (Hyperbolic and Contrastive Embedding using Decoupling with Baby Learning), en ny FSOD-model, der udnytter disse ideer gennem et hyperbolsk klassifikationshoved og et kontrastivt hoved. For at håndtere den øgede kompleksitet og overgangen i transfer learning anvender vi 'Baby Learning': en gradvis træningsstrategi, der først tilpasser modellen til den nye opgave og derefter øger kompleksiteten. Modellen bygger på Faster R-CNN og afkobler moduler for at mindske konflikter mellem den klasseagnostiske region proposal network (RPN) og det klasseafhængige RCNN-hoved, som deler samme backbone, samt mellem lokalisering og klassifikation. Vores eksperimeter viser, at HyCo-DeB overgår den hidtidige state-of-the-art på Pascal VOC og er på niveau med state-of-the-art på MS COCO.
Detecting new objects from only a few labeled examples—known as few-shot object detection (FSOD)—is highly desirable but remains difficult. A key factor is how visual features are represented as embeddings. Many systems embed in Euclidean space, even though hyperbolic space can capture hierarchical information more naturally. Another way to improve embeddings is contrastive learning, which pulls same-class examples together and pushes different classes apart. We introduce HyCo-DeB (Hyperbolic and Contrastive Embedding using Decoupling with Baby Learning), a new FSOD model that applies a hyperbolic classification head and a contrastive head to optimize embeddings. To manage the added complexity and the shift in transfer learning, we use 'Baby Learning': a gradual training strategy that first adapts the model to the new task and then increases complexity. Built on Faster R-CNN, our model decouples modules to reduce conflicts between the class-agnostic region proposal network (RPN) and the class-aware RCNN head that share the same backbone, and between localization and classification. Experiments show that HyCo-DeB surpasses the previous state of the art on the Pascal VOC benchmark and matches state-of-the-art performance on MS COCO.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Documents