EndoAbS dataset: Endoscopic abdominal stereo image dataset for benchmarking 3D stereo reconstruction algorithms

Background: 3D reconstruction algorithms are of fundamental importance for augmented reality applications in computer-assisted surgery. However, few datasets of endoscopic stereo images with associated 3D surface references are currently openly available, preventing the proper validation of such algorithms. This work presents a new and rich dataset of endoscopic stereo images (EndoAbS dataset). Methods: The dataset includes (i) endoscopic stereo images of phantom abdominal organs, (ii) a 3D organ surface reference (RF) generated with a laser scanner and (iii) camera calibration parameters. A detailed description of the generation of the phantom and the camera-laser calibration method is also provided. Results: An estimation of the overall error in creation of the dataset is reported (camera-laser calibration error 0.43mm) and the performance of a 3D reconstruction algorithm is evaluated using EndoAbS, resulting in an accuracy error in accordance with state-of-the-art results (<2mm). Conclusions: The EndoAbS dataset contributes to an increase the number and variety of openly available datasets of surgical stereo images, including a highly accurate RF and different surgical conditions.