SAR-RARP50

Surgical instrument segmentation and action recognition on 50 robot-assisted radical prostatectomy (RARP) procedures — the MICCAI/EndoVis 2022 SAR-RARP50 challenge dataset. Video covers the dorsal venous complex (DVC) suturing phase, captured from the left channel of a da Vinci stereo endoscope.

Contents

One row per annotated frame: the decoded 1920x1080 RGB video frame and its instrument segmentation mask.

split	operations	archives	frames
train	40	44	13,043
test	10	10	3,252
total	50	54	16,295

Masks are annotated at 1 Hz from 60 fps video, which is exactly the rate the official challenge metric scores at, so this mirror is the complete segmentation benchmark. Both splits carry real ground truth. The raw video_left.avi files are not mirrored (31 GiB, ~98% unannotated); the 10 Hz action labels are kept in full, both as per-row columns and as action_annotations.parquet (162,705 rows).

Segmentation classes

Mutually exclusive; stored single-channel, values 0-9.

id	class	id	class
0	Background	5	Thread
1	Tool clasper	6	Suction tool
2	Tool wrist	7	Needle Holder
3	Tool shaft	8	Clamps
4	Suturing needle	9	Catheter

Action classes (0-7)

0 Other · 1 Picking-up the needle · 2 Positioning the needle tip · 3 Pushing the needle through the tissue · 4 Pulling the needle out of the tissue · 5 Tying a knot · 6 Cutting the suture · 7 Returning/dropping the needle

Columns

column	notes
image / mask	1920x1080; mask values 0-9
operation_id	1-50. Group on this — it collapses the two-part operations
video_id	source archive, e.g. video_11_1 (keeps part identity)
part	1 or 2 for operations 11/15/17/29, else null
frame_index	0-based index into the original video_left.avi
time_sec	frame_index / 60
action_label / action_name	10 Hz action annotation at this frame
challenge_subset	training_set_1 / training_set_2 / test
num_frames_video	total frames in the source video

Differences from the upstream release

This mirror is a faithful reformatting, with four upstream inconsistencies fixed. Each was verified against the actual bytes rather than the documentation.

1. Masks are stored single-channel. Upstream PNGs are 3-channel RGB with R==G==B despite the README calling them "grayscale". Channel 0 is kept — lossless, and ~2.7x smaller.
2. Action labels are 0-7. The upstream README table lists 1-8; it is off by one (its header row is copy-pasted from the segmentation table). The values in action_discrete.txt are 0-7, matching the paper, the challenge GestureList.txt, and the official evaluator's n_classes=8.
3. training_set_2 is corrected. The upstream README duplicates video 34 and omits video 14, so its two training lists cover 39 of 40 operations. Video 14 is restored, per the original challenge listing.
4. Operations 11/15/17/29 ship as two archives each whose frame numbering restarts at 0. video_id keeps the parts distinct while operation_id collapses them, so grouping on operation_id never splits one patient across folds — the challenge explicitly treats these as one operation.

Caveats

• Class imbalance is severe. Needle Holder appears in ~4% of frames and Suction tool in ~9%; pixel imbalance reaches ~1000:1. The official metric uses MeanIoU(ignore_empty=False), which awards 1.0 when a class is absent from both prediction and reference — so published scores (winner: 0.829 mIoU) are inflated relative to a mean-IoU-over-present-classes and are not directly comparable unless that convention is replicated.
• 201 masks (1.23%) are entirely background, 17 of them in test.
• Action class 5 ("Tying a knot") is effectively untestable: it appears in one test video only (video_46, 186 frames).
• Video is lossy MPEG-4 Part 2 / yuv420p; frames here are a lossless PNG encoding of that already-compressed source.

Overlap with other datasets

No patient overlap with EndoVis 2017/2018 (both porcine) or with CholecSeg8k / m2caiSeg (cholecystectomy). Shares patients with RARP45 (never publicly released) and with the SEDMamba surgical-error dataset (UCL RDR 10.5522/04/27992702), which re-annotates 48 of these same videos — do not benchmark against SEDMamba as if it were independent.

License

CC BY-NC-SA 4.0, inherited from both source deposits. ShareAlike applies: this reformatted derivative carries the same license. Non-commercial use only.

Note: third-party Kaggle mirrors of this dataset declare ODbL, which is incorrect and more permissive than the authors' terms.

Source

• Train: https://rdr.ucl.ac.uk/articles/dataset/SAR-RARP50_train_set/24932529 (DOI 10.5522/04/24932529)
• Test: https://rdr.ucl.ac.uk/articles/dataset/SAR-RARP50_test_set/24932499 (DOI 10.5522/04/24932499)
• Toolkit: https://github.com/surgical-vision/SAR_RARP50-evaluation

Citation

@article{psychogyios2024sarrarp50,
  title   = {SAR-RARP50: Segmentation of surgical instrumentation and Action
             Recognition on Robot-Assisted Radical Prostatectomy Procedures},
  author  = {Psychogyios, Dimitrios and Colleoni, Emanuele and Van Amsterdam,
             Beatrice and others},
  journal = {arXiv preprint arXiv:2401.00496},
  year    = {2024},
  doi     = {10.48550/arXiv.2401.00496}
}