RT-less: a multi-scene RGB dataset for 6D pose estimation of reflective texture-less objects

The 6D (6 Degree of freedom) pose estimation (or pose measurement) of machined reflective texture-less objects, which are common in industry, is a significant but challenging technique. It has attracted increasing attention in academia and industry. However, it is difficult to obtain suitable public datasets of such objects, which makes relevant studies inconvenient. Thus, we proposed the Reflective Texture-Less (RT-Less) object dataset, which is a new public dataset of reflective texture-less metal parts for pose estimation research. The dataset contains 38 machined texture-less reflective metal parts in total. Different parts demonstrate the symmetry and similarity of shape and size. The dataset contains 289 K RGB images and the same number of masks, including 25,080 real images, 250,800 synthetic images in the training set, and 13,312 real images captured in 32 different scenes in the test set. The dataset also provides accurate ground truth poses, bounding-box annotations and masks for these images, which makes RT-Less suitable for object detection and instance segmentation. To improve the accuracy of the ground truth, an iterative pose optimization method using only RGB images is proposed. Baselines of the state-of-the-art pose estimation methods are provided for further comparative studies. The dataset and results of baselines are available at: http://www.zju-rtl.cn/RT-Less/.