[Paper Review] Products-10K: A Large-scale Product Recognition Dataset

Yalong Bai, Yuxiang Chen+3 more|arXiv (Cornell University)|Aug 24, 2020

Advanced Image and Video Retrieval Techniques10 references32 citations

TL;DR

Introduces Products-10K, a large-scale, human-labeled SKU-level product dataset with ~150k images across 10k SKUs, plus practical fine-grained recognition tricks. Provides experiments showing high-resolution training and metric-aligned losses improve accuracy.

ABSTRACT

With the rapid development of electronic commerce, the way of shopping has experienced a revolutionary evolution. To fully meet customers' massive and diverse online shopping needs with quick response, the retailing AI system needs to automatically recognize products from images and videos at the stock-keeping unit (SKU) level with high accuracy. However, product recognition is still a challenging task, since many of SKU-level products are fine-grained and visually similar by a rough glimpse. Although there are already some products benchmarks available, these datasets are either too small (limited number of products) or noisy-labeled (lack of human labeling). In this paper, we construct a human-labeled product image dataset named "Products-10K", which contains 10,000 fine-grained SKU-level products frequently bought by online customers in JD.com. Based on our new database, we also introduced several useful tips and tricks for fine-grained product recognition. The products-10K dataset is available via https://products-10k.github.io/.

Motivation & Objective

• Motivate the need for a large-scale, fine-grained SKU-level product dataset for e-commerce recognition.
• Create a human-labeled dataset covering diverse product categories with low label noise.
• Provide practical tips and tricks to improve fine-grained product recognition without extra inference cost.

Proposed method

• Assemble a large, human-labeled SKU-level product dataset (Products-10K) with in-shop and customer images.
• Ensure low label noise through multi-expert verification (each image checked by at least three experts).
• Propose training tricks: high-resolution input, balanced-subset finetuning, and metric-guided loss design to align optimization with accuracy.
• Demonstrate that high-resolution training yields significant accuracy gains without changing inference cost.
• Introduce an accuracy-focused loss (L_acc) to better optimize for top-1 accuracy during finetuning.

Experimental results

Research questions

• RQ1 Can a large-scale, fully human-labeled SKU-level dataset improve fine-grained product recognition?
• RQ2 Do high-resolution inputs, balanced fine-tuning, and metric-aligned losses improve top-1 accuracy on a complex product dataset?
• RQ3 How do in-shop and customer images impact generalization and noise robustness in SKU recognition?

Key findings

• Products-10K contains nearly 10k SKUs and about 150k images with noise rate below 0.5%.
• High-resolution training (512→448 crop) yields 4.9 percentage points in top-1 accuracy over 224×224 baselines.
• Balanced-subset finetuning provides a further gain (approx. 1.8% when combined with other tricks).
• Metric-guided loss improves optimization for accuracy without added inference cost, achieving the best reported top-1 on the validation set.
• A single EfficientNet-B3 model achieves 64.12% top-1 accuracy with 448×448 input and metric loss finetune.