Garbage object detection is of central importance to achieve accurate garbage sorting. In this work, we propose an improved object detection algorithm for sorting complex garbage objects on the assembly line. This algorithm utilizes the Faster RCNN as the baseline and improves its performance with several tips, including deformable convolution, PAFPN, mixup, cutout, color jitter, multi-scale training and ImageNet transfer learning. Considering the lack of public datasets for the garbage object detection task, we synthesize a highly realistic multi-object garbage detection dataset, M-GarbageNet, with 20000 training images, 5000 validation images and 5000 test images. Finally, we evaluate our algorithm than Faster RCNN, and obtain a 8.2% relative improvement on the M-GarbageNet dataset.