[Paper Review] Semi-Supervised Recognition under a Noisy and Fine-grained Dataset

Cheng Cui, Zhi Ye+8 more|arXiv (Cornell University)|Jun 18, 2020

Advanced Neural Network Applications9 references4 citations

TL;DR

This paper presents a semi-supervised approach for fine-grained bird recognition under noisy and limited-supervised conditions, combining pseudo-labeling with hybrid generic and fine-grained recognition models. Using PaddleClas for model training and leveraging unlabeled data via pseudo-label mining, the method achieved third place in the FGVC7 challenge, demonstrating effectiveness in handling subtle visual differences and data noise.

ABSTRACT

Simi-Supervised Recognition Challenge-FGVC7 is a challenging fine-grained recognition competition. One of the difficulties of this competition is how to use unlabeled data. We adopted pseudo-tag data mining to increase the amount of training data. The other one is how to identify similar birds with a very small difference, especially those have a relatively tiny main-body in examples. We combined generic image recognition and fine-grained image recognition method to solve the problem. All generic image recognition models were training using PaddleClas . Using the combination of two different ways of deep recognition models, we finally won the third place in the competition.

Motivation & Objective

• Address the challenge of fine-grained bird recognition with minimal labeled data and high noise levels.
• Improve model generalization by effectively utilizing large-scale unlabeled data in a semi-supervised setting.
• Overcome difficulties in distinguishing visually similar bird species with small inter-class variations.
• Integrate generic image recognition and fine-grained recognition techniques to enhance feature discrimination.
• Achieve high performance in a real-world competition setting with limited annotations and noisy data.

Proposed method

• Employed pseudo-labeling to mine and incorporate unlabeled data into the training process, increasing effective training data size.
• Combined pre-trained generic image recognition models with fine-grained recognition architectures to enhance feature learning.
• Trained all generic models using PaddleClas, a deep learning library for computer vision.
• Used a dual-stream model design to jointly optimize for both general and fine-grained visual characteristics.
• Applied data augmentation and self-training strategies to improve robustness to noisy examples.
• Leveraged model confidence scores to filter and select high-quality pseudo-labels during training.

Experimental results

Research questions

• RQ1 How can unlabeled data be effectively leveraged in a semi-supervised setting for fine-grained recognition?
• RQ2 What is the impact of combining generic and fine-grained recognition models on performance under noisy data?
• RQ3 How can small inter-class variations in bird species be reliably distinguished using deep learning?
• RQ4 What role does pseudo-label quality play in semi-supervised fine-grained classification?
• RQ5 Can hybrid model architectures outperform single-architecture models in low-supervision, high-noise scenarios?

Key findings

• The proposed method achieved third place in the FGVC7 competition, demonstrating strong performance under challenging conditions.
• Pseudo-labeling significantly improved model accuracy by expanding the effective training set with high-confidence predictions.
• The hybrid model architecture combining generic and fine-grained recognition components outperformed single-architecture baselines.
• The use of PaddleClas enabled efficient and stable training of generic models, contributing to overall performance.
• The method showed robustness to noisy annotations and small visual differences between bird species.
• High-confidence pseudo-labels were critical in maintaining performance gains without introducing error propagation.