Built independently by an author, for readers. Read the story and support ChapterPal
ChapterPal Logo
Log inSign up

keyword

object knowledge distillation

Object knowledge distillation is a machine learning technique in which fine-grained, object-level representations, such as spatial boundaries, localized features, and category semantics from a teacher model or dedicated detector, are transferred to guide the training of a student neural network. Unlike standard distillation approaches that transfer global, whole-image predictions or embeddings, object knowledge distillation focuses on localized visual concepts within specific regions of an image. This technique is commonly applied in object detection and multimodal vision-language modeling to enhance region-level semantic understanding, facilitate alignment between localized visual features and descriptive text, and produce efficient models that can recognize fine-grained concepts without requiring computationally expensive object-extraction pipelines during inference.

1 item

Multi-Grained Vision Language Pre-Training: Aligning Texts with Visual Concepts

Multi-Grained Vision Language Pre-Training: Aligning Texts with Visual Concepts

Yan Zeng, Xinsong Zhang, Hang Li

OrganizationsByteDance

Why you should read this

Proposes X-VLM, a vision-language pre-training framework that learns multi-grained cross-modal alignments across objects, regions, and full images without relying on object detectors, achieving state-of-the-art results on visual grounding, retrieval, and reasoning tasks.

Most existing methods in vision language pre-training rely on object-centric features extracted through object detection and make fine-grained alignments between the extracted features and texts. It is challenging for these methods to learn relations among multiple objects. To this end, we propose a new method called X-VLM 1 to perform ‘multi-grained vision language pre-training.’ The key to learning multi-grained alignments is to locate visual concepts in the image given the associated texts, and in the meantime align the texts with the visual concepts, where the alignments are in multi-granularity. Experimental results show that X-VLM effectively leverages the learned multi-grained alignments to many downstream vision language tasks and consistently outperforms state-of-the-art methods.

Added

2026-09-26