MicroDepth: a stereo-depth dataset for microsurgical scenes of soft tissues

MicroDepth contains 3,586 synchronized stereo pairs (1280 × 960) with registered structured-light depth, acquired in 34 sessions of static, ex vivo microsurgical scenes. The six acquisition categories are BrainTissue, MuscleTissue, OcularTissue, RobotEffector, CoreInst and AuxInst; the tissues are porcine brain, porcine eye and chicken muscle. Industrial RGB cameras (stereo baseline 31.51 mm) and a PhoXi XS structured-light scanner were used in sequential acquisition at a working distance of about 210–250 mm. The scanner is translated along one calibrated stage axis to positions equivalent to the camera viewpoints (STEPS). Each frame provides rectified left and right RGB images, left and right depth arrays, a legacy left disparity array, a published left validity mask, a left–right consistency score and metadata. The depth arrays hold original-camera axial Z on rectified image grids; the accompanying code converts them to the rectified pixel-centre axial convention used in the descriptor. The consistency score is computed from left–right depth agreement and is not the scanner's native confidence. Version 1.1.0 adds a README, a session-disjoint partition (23/5/6 sessions; 2,656/405/525 frames for training, validation and test; seed 20261025), a validation directory with the region definitions and masks, fixed analysis settings, category-level results and simulation outputs, and the stage-axis calibration and acquisition settings. The data archive MicroDepth.zip is unchanged from version 1.0.0. Repeatability and axial depth-difference error were measured on a certified gauge-block target: the unfiltered left-view 95th-percentile per-pixel standard deviation over ten acquisitions is 10.6 µm, and the mean absolute axial depth-difference error on two nominal 10 mm steps is 0.1025 mm. These values describe the stated target configuration and are not an absolute accuracy for tissue. The partition separates acquisition sessions and does not certify specimen independence. Companion paper: "A stereo-depth dataset for microsurgical scenes of soft tissues" (Scientific Data, under revision). Processing and evaluation code: https://github.com/afpapqy/MicroDepth-code.