Data

Training Data Overview

Our training dataset is made up of 16 robotic nephrectomy procedures recorded using da Vinci Xi systems in porcine labs. The original video data was recorded at 60 Hz and to reduce labelling cost we subsample this to 2 Hz. Sequences with little or no motion are manually removed to leave 149 frames per procedure. Video frames are 1280x1024 and we provide the left and right eye camera image as well as the stereo camera calibration parameters. Labels are only provided for the left image. 

In each frame we hand label several man-made and anatomical objects. The annotations were performed by hand by several Intuitive Surgical employees with knowledge of porcine anatomy. Only a single annotator worked on each dataset. Best efforts will be made to ensure anatomical correctness of all labels however we cannot guarantee that some errors do not occur, particularly in cases where tissue type is ambiguous. The classes found in the training and test will be:

  • da Vinci robotic surgical instrument parts
    • Shaft 
    • Wrist 
    • Jaws
  • Drop in Ultrasound Probe
  • Suturing Needles
  • Suturing thread
  • Clips/clamps
  • Kidney parenchyma
    • Fascia covered
    • Uncovered
  • Small bowel
  • Background tissue
  • Each class will have a distinct numerical label in a ground truth image. A supplied json file will contain the class name to numerical label mapping.

Test Data Overview

We will release 2x500 frame sequences as a test sequence. They will be from different porcine procedures as the training data with no instrument types or objects that are unseen in the training data. Similar to the training data, the test procedures will consist of porcine nephrectomies performed on the da Vinci Xi robot. 

Evaluation Criteria

The challenge will be ranked on the mean intersection over union (IoU) metric. This means we compute the IoU for each class which is present in a frame and then average over these scores for a per-frame score. 

Participant conditions

  • All methods used to generate the results must be fully automatic.
  • Participants may use publicly available datasets to augment or pre-train networks. However, this usage must be made explicit in any description of the method. 
  • Only a single submission will be considered for each team. If multiple submissions are made, the most recent will be used unless requested by the participating team. The only exception to this rule is if the team is approved to submit multiple different methods. Permission for this type of submission should be made by request to the challenge organizers.