SpectralWaste Dataset: Multimodal Data for Waste Sorting Automation

TL;DR

SpectralWaste is presented, the first dataset collected from an operational plastic waste sorting facility that provides synchronized hyperspectral and conventional RGB images and demonstrates how hyperspectral imaging can bring a boost to RGB-only perception in these realistic industrial settings without much computational overhead.

Abstract

The increase in non-biodegradable waste is a worldwide concern. Recycling facilities play a crucial role, but their automation is hindered by the complex characteristics of waste recycling lines like clutter or object deformation. In addition, the lack of publicly available labeled data for these environments makes developing robust perception systems challenging. Our work explores the benefits of multimodal perception for object segmentation in real waste management scenarios. First, we present SpectralWaste, the first dataset collected from an operational plastic waste sorting facility that provides synchronized hyperspectral and conventional RGB images. This dataset contains labels for several categories of objects that commonly appear in sorting plants and need to be detected and separated from the main trash flow for several reasons, such as security in the management line or reuse. Additionally, we propose a pipeline employing different object segmentation architectures and evaluate the alternatives on our dataset, conducting an extensive analysis for both multimodal and unimodal alternatives. Our evaluation pays special attention to efficiency and suitability for real-time processing and demonstrates how HSI can bring a boost to RGB-only perception in these realistic industrial settings without much computational overhead.

Figures (6)

• Figure 1: The multimodal setup in the waste sorting facility contains two synchronized line-scan cameras that gather RGB and hyperspectral data. Left: Real prototype installed in the facility. Middle: Diagram of the setup for data capture. Both hyperspectral and RGB cameras are housed in an industrial enclosure and completed with artificial lighting to ensure correct image capture. Right: Example of a scene as captured by the RGB camera and by the hyperspectral camera (with a false-color for visualization purposes).
• Figure 2: Examples of images included in the dataset. The first row shows images captured by the RGB camera, the second row shows the ground-truth RGB annotations and the third row shows the same scenes captured by the hyperspectral camera, annotated with labels obtained using the proposed label-transfer algorithm. For visualization purposes, we manually select three out of all the hyperspectral bands.
• Figure 3: Pipeline of the segmentation process. Path 1 represents the data flow for the initial architectures designed for unimodal RGB segmentation (MiniNet-v2 and SegFormer). The dashed path 2 represents the HSI data flow in the multimodal CMX architecture.
• Figure 4: Visualization of the main steps of the proposed label transfer method: (a) initial mask superposed to HSI; (b) original manually annotated mask (in pink) superposed to RGB image; (c) contours of the extracted connected components; (d) uniformly-sampled points (red); (e) outermost points (green); (f) COTR matching for selected points in RGB (top) to the HSI image (bottom) used to compute an affine transformation for the mask; (g) resulting mask after projection into the HSI image.
• Figure 5: Qualitative results of the label transfer evaluation. The manually annotated masks are shown in green and the resulting mask from each method is in red. First row visualizes the proposed label transfer (LT) and second row the manual alignment (MA).