Due to the current labor shortage, automation by robots has been expected in society. In manual sorting of garbage in recycling factories, there is a risk of injury due to sharp garbage, and robots are needed to replace them. In this paper, we propose an improved method of garbage sorting using thermal images. Previously, we classified three types of beverage container garbage from thermal images, but it could not cope with dense garbage. In this work, material classification is performed for each pixel of the thermal image, followed by clustering to correctly separate and classify garbage, even dense garbage. In the experiment, we collected thermal images of heated garbage on hot conveyor for classification, and verified the accuracy of the computed material and object maps, and compared them with previous work.