Precision at Scale: Domain-Specific Datasets On-Demand

TL;DR

The paper tackles the problem that self-supervised pretraining on large general-domain datasets may underperform domain-specific pretraining. It introduces Precision at Scale (PaS), a modular, autonomous pipeline that constructs on-demand domain-specific datasets by combining in-domain concept discovery with real and synthetic image collection followed by automated curation. Across bird and food domains, PaS-generated datasets reach rich domain coverage and yield superior pretraining gains compared to ImageNet baselines, including scenarios where PaS-F outperforms ImageNet-21K while being far smaller. This work demonstrates that high-quality, domain-focused data can outperform large general datasets for pretraining, offering a practical path to scalable, domain-aware SSL with flexible dataset sizing.

Abstract

In the realm of self-supervised learning (SSL), conventional wisdom has gravitated towards the utility of massive, general domain datasets for pretraining robust backbones. In this paper, we challenge this idea by exploring if it is possible to bridge the scale between general-domain datasets and (traditionally smaller) domain-specific datasets to reduce the current performance gap. More specifically, we propose Precision at Scale (PaS), a novel method for the autonomous creation of domain-specific datasets on-demand. The modularity of the PaS pipeline enables leveraging state-of-the-art foundational and generative models to create a collection of images of any given size belonging to any given domain with minimal human intervention. Extensive analysis in two complex domains, proves the superiority of PaS datasets over existing traditional domain-specific datasets in terms of diversity, scale, and effectiveness in training visual transformers and convolutional neural networks. Most notably, we prove that automatically generated domain-specific datasets lead to better pretraining than large-scale supervised datasets such as ImageNet-1k and ImageNet-21k. Concretely, models trained on domain-specific datasets constructed by PaS pipeline, beat ImageNet-1k pretrained backbones by at least 12% in all the considered domains and classification tasks and lead to better food domain performance than supervised ImageNet-21k pretrain while being 12 times smaller. Code repository: https://github.com/jesusmolrdv/Precision-at-Scale/

Figures (15)

• Figure 1: Stage 1 workflow. Based on the output concepts of an initial prompt, we extend the output by chaining N number of prompts. Once we saturate the diversity, we filter them using by prompting an auxiliary LLM.
• Figure 2: Stage 2 Workflow. For every valid concept extracted on the first stage, we collect the most N similar images from a real-data source. Similarly, we prompt an image generation algorithm using the concept to produce a set of synthetic images. The combination of both sets form the unfiltered version of the desired dataset.
• Figure 3: Comparative analysis of lexical concept distributions in bird domain.
• Figure 4: Comparative analysis of lexical concept distributions in food domain.
• Figure 5: Comparison of image embeddings distributions across Bird and Food domains.