[论文解读] TACO: Trash Annotations in Context for Litter Detection

Pedro F. Proença, Pedro Simões|arXiv (Cornell University)|Mar 16, 2020

Microplastics and Plastic Pollution参考文献 7被引用 112

一句话总结

TACO 是一个用于垃圾检测和分割的开放图像数据集，通过众包创建，使用 Mask R-CNN 在两个分类法（TACO-1 和 TACO-10）进行4折交叉验证评估；结果有希望但指出需要更多注释与对微小对象的高分辨率处理。

ABSTRACT

TACO is an open image dataset for litter detection and segmentation, which is growing through crowdsourcing. Firstly, this paper describes this dataset and the tools developed to support it. Secondly, we report instance segmentation performance using Mask R-CNN on the current version of TACO. Despite its small size (1500 images and 4784 annotations), our results are promising on this challenging problem. However, to achieve satisfactory trash detection in the wild for deployment, TACO still needs much more manual annotations. These can be contributed using: http://tacodataset.org/

研究动机与目标

• 激发创建一个情境丰富的垃圾数据集，以改善野外的自主垃圾检测。
• 描述 TACO 数据集、它的注释分类法以及众包工具。
• 评估 Mask R-CNN 在垃圾检测与分割方面在两种分类法上的性能。
• 讨论用于平衡背景并提升训练的数据增强与移植方法。
• 识别局限性并概述扩展注释与改进小对象检测的未来方向。

提出的方法

• 介绍 TACO，包含 1500 张高分辨率图像和 4784 条来自众包的注释。
• 用一个分层的分类法对垃圾进行标注，涵盖 60 个类别，跨 28 个上位类别，并再加一个未标注垃圾类别。
• 在两个任务上评估 Mask R-CNN（ResNet-50 配 FPN，输入尺寸 1024x1024）：TACO-1（无类别垃圾）和 TACO-10（10 种垃圾类别）。
• 使用4折交叉验证（80/10/10 拆分）以及基于 IoU 阈值的平均精度（AP）作为评估指标。
• 实验 Mask R-CNN 的三种排序分数：class_score、litter_score、ratio_score，以优化 AP。
• 应用数据增强（高斯模糊、AWG 噪声、曝光/对比度变换、旋转）并将 320 个标注实例移植到 Flickr 图像中以增广训练。

实验结果

研究问题

• RQ1 Can Mask R-CNN effectively detect and segment litter in diverse, real-world environments using the TACO dataset?
• RQ2 How does the choice of ranking score (class_score, litter_score, ratio_score) affect AP for litter detection and classification?
• RQ3 What is the impact of background variety and object size on detection performance, especially for tiny objects like cigarettes?
• RQ4 Does expanding the taxonomy from a single litter class (TACO-1) to multiple classes (TACO-10) improve discriminative performance or introduce confusion?
• RQ5 How can crowd-sourced annotations and tranplantation techniques improve training data for litter detection in the wild?

主要发现

• AP results depend on the scoring strategy; ratio_score improves AP for TACO-10 and does not reduce AP for TACO-1.
• Tiny objects (e.g., cigarettes) significantly reduce detection performance due to small bounding boxes after resizing.
• Cans and bottles are detected more reliably than cigarettes, with some confusion between Plastic bag and Other categories.
• Transplanting segmentations with soft masks via distance transforms reduces edge artifacts compared to hard transplantation.
• The dataset’s performance is promising but indicates substantial room for improvement with more annotations and higher input resolution.
• Some success is shown in handling transparent objects, but background diversity (e.g., ocean waves) still challenges generalization.