[论文解读] iCassava 2019 Fine-Grained Visual Categorization Challenge

Ernest Mwebaze, Timnit Gebru+3 more|arXiv (Cornell University)|Aug 8, 2019

Smart Agriculture and AI参考文献 11被引用 54

一句话总结

本文提出一个木薯叶数据集，包含带标签和未标签图像，用于 Kaggle 挑战，促进在适合移动设备的模型上进行半监督的细粒度病害分类。它报告数据集统计、挑战设置以及使用 ResNet 架构的基线结果。

ABSTRACT

Viral diseases are major sources of poor yields for cassava, the 2nd largest provider of carbohydrates in Africa.At least 80% of small-holder farmer households in Sub-Saharan Africa grow cassava. Since many of these farmers have smart phones, they can easily obtain photos of dis-eased and healthy cassava leaves in their farms, allowing the opportunity to use computer vision techniques to monitor the disease type and severity and increase yields. How-ever, annotating these images is extremely difficult as ex-perts who are able to distinguish between highly similar dis-eases need to be employed. We provide a dataset of labeled and unlabeled cassava leaves and formulate a Kaggle challenge to encourage participants to improve the performance of their algorithms using semi-supervised approaches. This paper describes our dataset and challenge which is part of the Fine-Grained Visual Categorization workshop at CVPR2019.

研究动机与目标

• Motivate automated cassava disease diagnosis to support smallholder farmers in Africa.
• Provide a realistic, in-situ image dataset with labeled and unlabeled data reflecting field challenges.
• Encourage semi-supervised and lightweight models suitable for mobile deployment.
• Promote fair evaluation via a Kaggle challenge with public and private leaderboards.

提出的方法

• Assemble and annotate a dataset of cassava leaf images from ~200 farmers in Uganda.
• Label each image with one primary disease class among five categories (Healthy, CMD, CBSD, CBB, CGM).
• Include a large unlabeled image set to foster semi-supervised learning approaches.
• Release data processing steps including cropping and consistency checks to reflect real-world field conditions.
• Employ ResNet-based architectures with data augmentation as the baseline approach.
• Use overall accuracy as the evaluation metric on public and private Kaggle leaderboards.

实验结果

研究问题

• RQ1 Can semi-supervised learning improve fine-grained cassava disease classification with abundant unlabeled data?
• RQ2 How well do models handle in-field imaging challenges such as varying backgrounds, multiple diseases per plant, and imperfect focus?
• RQ3 What accuracy levels are achievable by lightweight models suitable for smartphones in real-world deployment?

主要发现

• Dataset comprises 9,436 labeled and 12,595 unlabeled cassava leaf images.
• Five classes: Healthy, CMD, CBSD, CBB, CGM with specified train/test splits.
• Top solutions achieved around 93% accuracy on the private leaderboard.
• Winning approach leveraged unlabeled data to gain about 1% improvement.
• All three top solutions used the ResNet architecture with data augmentation.
• Challenge demonstrates feasibility of semi-supervised, fine-grained classification in agricultural settings.