[Paper Review] PP-ShiTu: A Practical Lightweight Image Recognition System

Shengyu Wei, Ruoyu Guo+11 more|arXiv (Cornell University)|Nov 1, 2021

Advanced Image and Video Retrieval Techniques3 references6 citations

TL;DR

PP-ShiTu is a practical, lightweight image recognition system that integrates mainbody detection, feature extraction, and vector search into a unified pipeline. By leveraging techniques like metric learning, knowledge distillation, model quantization, and DeepHash, it achieves high accuracy and inference speed across diverse image recognition tasks, outperforming existing models in speed and efficiency while maintaining competitive performance on multiple benchmarks.

ABSTRACT

In recent years, image recognition applications have developed rapidly. A large number of studies and techniques have emerged in different fields, such as face recognition, pedestrian and vehicle re-identification, landmark retrieval, and product recognition. In this paper, we propose a practical lightweight image recognition system, named PP-ShiTu, consisting of the following 3 modules, mainbody detection, feature extraction and vector search. We introduce popular strategies including metric learning, deep hash, knowledge distillation and model quantization to improve accuracy and inference speed. With strategies above, PP-ShiTu works well in different scenarios with a set of models trained on a mixed dataset. Experiments on different datasets and benchmarks show that the system is widely effective in different domains of image recognition. All the above mentioned models are open-sourced and the code is available in the GitHub repository PaddleClas on PaddlePaddle.

Motivation & Objective

• To address the inefficiency and high migration cost of adapting image recognition models across different applications like face recognition, re-identification, and product retrieval.
• To design a unified, practical pipeline that generalizes across multiple image recognition domains with minimal retraining.
• To optimize model speed and accuracy using lightweight architectures, distillation, quantization, and hashing techniques.
• To enable efficient inference on CPU and mobile devices through model compression and optimized inference pipelines.
• To provide open-source, reusable models and code for broad adoption in real-world applications.

Proposed method

• Uses PP-PicoDet with PP-LCNet backbone for efficient mainbody detection in images.
• Employs metric learning with ArcMargin loss to enhance feature discrimination in the embedding space.
• Applies knowledge distillation (U-DML and SSLD) to transfer knowledge from larger models to lightweight student networks.
• Utilizes model quantization to reduce model size and accelerate inference on CPU.
• Implements DeepHash to compress feature vectors into binary codes for faster vector search.
• Combines all modules into an end-to-end pipeline: detect main objects → extract features → search nearest neighbors in a vector database.

Experimental results

Research questions

• RQ1 Can a unified, lightweight pipeline effectively handle diverse image recognition tasks such as product, face, and vehicle recognition?
• RQ2 How do knowledge distillation and metric learning jointly improve accuracy in lightweight models?
• RQ3 To what extent does feature quantization via DeepHash reduce storage and inference time without sacrificing retrieval accuracy?
• RQ4 How does the integration of PP-PicoDet and PP-LCNet balance speed and accuracy on CPU-optimized inference?
• RQ5 What is the performance gain of using hybrid training data across multiple datasets for generalization?

Key findings

• PP-ShiTu achieves a 194ms average latency on Intel Xeon Gold 6148 CPU, reducing inference time by over 80% compared to a ResNet-50-based server model.
• The system improves recall@1 by 0.7% on average across six datasets using U-DML knowledge distillation, with no extra inference cost.
• The binary feature model using DeepHash reduces index size from 285MB to 8.9MB and retrieval time from 117ms to 21ms, while maintaining 78% recall@1 on the Ali-Product dataset.
• PP-LCNet-1x achieves 71.32% top-1 accuracy with only 3.0M parameters and 2.46ms latency on Intel Xeon, outperforming MobileNetV3 and ShuffleNetV2 in speed-accuracy trade-off.
• The system maintains consistent performance across diverse datasets including iCartoonFace, VeRI-Wild, and SOP, with average recall@1 of 0.859 using PP-LCNet-2.5x with U-DML.
• Ablation studies confirm that combining metric learning, distillation, and quantization yields synergistic improvements in accuracy and efficiency.