FastLabel needed to prepare large computer-vision datasets for autonomous-driving and industrial workloads, but manual captioning and duplicate review made image curation slow, expensive, and too error-prone to keep up with customer demand.
Reported impact
FastLabel reduced a 10,000-image captioning effort from 333 hours to 14.6 hours, ran text embedding and semantic deduplication in minutes, lowered the total cost of processing 10,000 images to under $57, and used the pipeline to launch its commercial FastLabel Data Curation service.
FastLabel built an image-data curation service on NVIDIA NeMo Curator running on NVIDIA A100 GPUs in Google Cloud Platform. The workflow starts with large autonomous-driving image sets and generates detailed domain-specific captions through vision-language models hosted on GCP. Those captions are embedded and then processed through NeMo Curator's semantic deduplication pipeline so the service can identify visually redundant examples that rule-based filters would miss. By turning image semantics into text, FastLabel made it practical to curate long-tail datasets for perception systems while drastically reducing manual review. The same pipeline now underpins FastLabel Data Curation, a productized service that produces automatically tagged, deduplicated image datasets faster, at lower cost, and with less wasted training compute for customers building autonomous and industrial vision models.
Raw image datasets are captioned by vision-language models on GCP, captions are embedded and passed into NeMo Curator on A100 GPUs, semantically similar examples are removed, and the cleaned datasets are delivered through FastLabel's data-curation service for downstream autonomous-vision training.
Captioning of 10,000 images in approximately 14.6 hours.
FastLabel implemented the NeMo Curator image processing features on NVIDIA A100 GPUs on Google Cloud Platform.
For FastLabel, the solution enabled the launch of its FastLabel Data Curation service, providing customers with rapid, reliable access to high-quality, automatically tagged, and deduplicated datasets.
The core semantic deduplication step costing only $0.26 using an A100 GPU on Google Cloud.
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
The Organization JSON-LD lists addressLocality '新宿区', and addressCountry 'JP', and numberOfEmployees value 44.
Related
NVIDIA Sample Sep 22, 2026 Sluicebox Sluicebox needed to turn fragmented bills of materials, supplier records, PDFs, CAD files, and emissions inventories into timely product carbon assessments, but manual life-cycle-analysis work was slow, expensive, and… 51-80% lower inference costs Lucy · NVIDIA Nemotron 3 Ultra · carbon graph NVIDIA Sample Sep 15, 2026 Lambda Lambda needed to raise token throughput and improve AI-cloud economics, but fixed rack and facility power budgets were leaving GPU capacity stranded and limiting how many nodes could run safely at once. +17% Hybrid Training Throughput Dynamic Power Software · NVIDIA DSX AI Factory Platform · NVIDIA DSX MaxLPS NVIDIA Sample Sep 1, 2026 Onetera Onetera needed to help more than 90 municipalities turn fragmented regulations, zoning maps, permit rules, and legacy documents into fast resident-facing services without imposing heavy IT overhead or expensive manual… 1.7x increase in data throughput GPT OSS 20B · Government Intelligence Agent · LangGraph 