Object-Centric Image Resampling for Rare-Category Neural Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image resampling methods fail to effectively address the challenges of long-tailed datasets in object detection and localization, where images often contain a mixture of rare and frequent objects, leading to imbalanced training and biased performance on rare categories.
Innovation Solution
Introduce Object-Centric Sampling (OCS) with a dynamic memory bank to augment object-level features and bounding boxes, combining with image resampling to create a unified resampling framework that efficiently samples targeted categories, using a memory bank to store and retrieve object-level features for episodic augmentation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional image resampling methods are used to balance training data, then the amount of training data for rare categories is increased, but the performance on rare categories remains biased due to mixture with frequent objects in images
Solution Approach 1:
The patent segments images into individual object instances and creates separate training samples for each object. Instead of treating entire images as training units, the system extracts and resamples individual rare objects independently, allowing rare category objects to be trained separately from frequent objects that appear in the same images. This segmentation resolves the contamination effect where frequent objects dominated the training signal.
Solution Approach 2:
The patent extracts specific object-level features (bounding boxes, segmentation masks, class labels) from images and isolates rare category objects from their original image context. By extracting only the relevant rare objects and their associated metadata, the system creates purified training samples that focus exclusively on rare categories without the confounding presence of frequent objects.
2Measurement precision
If object-centric sampling with memory bank is implemented to improve rare category detection, then detection accuracy on rare classes improves by 1.89%, but computational resources and system complexity increase
Solution Approach 1:
The patent implements a memory bank that pre-stores object-level features (bounding boxes, masks, labels) of rare category objects during a first training phase. This preliminary action allows the system to quickly retrieve and resample rare objects during subsequent training without performing complex image processing in real-time, reducing computational overhead while maintaining high detection accuracy.
Solution Approach 2:
The patent creates copies of rare object instances by retrieving stored object-level features from the memory bank and generating synthetic training samples. Instead of processing original images repeatedly, the system copies and reuses extracted object features to create diverse training variations, improving rare category detection while avoiding redundant computational work.
3Quantity of substance
If entire images are resampled to increase rare category representation, then more rare objects are included in training batches, but frequent objects in the same images dominate the training signal and dilute the rare category learning
Solution Approach 1:
The patent segments images into individual object instances and creates separate training samples for each object. Instead of treating entire images as training units, the system extracts and resamples individual rare objects independently, allowing rare category objects to be trained separately from frequent objects that appear in the same images. This segmentation resolves the contamination effect where frequent objects dominated the training signal.
Solution Approach 2:
The patent applies different treatment to different parts of the data: rare category objects are extracted and resampled with higher frequency, while frequent objects are either excluded or given lower weight. This local quality approach ensures that training resources are concentrated on rare categories where they are most needed, rather than uniformly processing all objects in images.
Data Source
AI summary
Apparatuses, systems, and techniques to modify a set of training data used for machine learning. In at least one embodiment, a set of images used for training a machine learning system is resampled by augmenting the set of images with additional images of under represented object types extracted from portions of existing training images in the set.