Allelx/iris-instrument-clearing

IRIS surgical instrument clearing order

96 photographs of five surgical instruments on a table, each with hand-corrected instrument masks and an exact record of which instrument is lying on which, plus an order they can be cleared in without ever lifting one out from under another.

The task: look at the tray, work out the stacking, and give a clearing order.

The point of the dataset is where the ground truth comes from. A robot arm laid every scene out itself, one instrument at a time, in a recorded order — so the layering is not an annotator's judgement about a photograph, it is a log of what the machine did. That is very hard to get from images alone, and it is why these 96 scenes are worth more than 96 hand-labelled ones.

Every mask here was corrected by hand in CVAT, so mask quality is uniform across all three splits.

What a model sees, and what it answers

Input — one image: the survey frame with each instrument outlined and labelled with its class name.

The classes are given rather than asked for. A detector already supplies them at runtime, so the model is left with the part a detector cannot do: working out what is lying across what.

Output — one JSON object, with the keys in this order (u0013_arr1-3-2-4-5):

{"contacts":[["artery_forceps","tissue_forceps"]],
 "free":[{"class":"scalpel_handle_large","reason":"isolated"},
         {"class":"scalpel_handle_small","reason":"isolated"},
         {"class":"surgical_scissors","reason":"isolated"},
         {"class":"tissue_forceps","reason":"on_top"}],
 "pick_order_classes":["scalpel_handle_small","tissue_forceps",
                       "surgical_scissors","artery_forceps","scalpel_handle_large"],
 "topmost_class":"scalpel_handle_small"}

contacts is [lower, upper] — the first instrument is the one underneath. reason is on_top when an instrument lies across something, isolated when it touches nothing at all.

The key order is part of the target

contacts comes before pick_order_classes deliberately. A model generates left to right and cannot revise, so a justification emitted after the answer is produced by a model that has already committed to that answer, and will agree with it whether or not it was right. Emitted before, the order is generated with the stacking already in context and has to stay consistent with it.

The pick order is not a fact about the scene; the contact graph is, and the order follows from it by a rule. So answer.strict_contact_json in each target is the exact completion text, with the key order frozen into the string. Parsing and re-serialising it will re-sort the keys and undo this.

Layout

input_data/<cycle>/rgb.png            the survey frame
input_data/<cycle>/rgb_annotated.png  the model input
input_data/<cycle>/masks.json         polygons per instrument, in pixels
input_data/<cycle>/motion.json        the arm's log for that cycle
input_data/<cycle>/video.json         frame times (the video is not included)
training_target/<cycle>.json          labels, both variants, and the completion
validation/…                          same two folders, held-out cycles
test/…                                same two folders — sealed, see below
manifest.json, split.csv

Validation and test hold their own inputs and targets, so a held-out cycle is absent from the training folders rather than merely marked.

Splits

Split	Cycles	Arrangement groups
train	61	16
validation	19	5
test	16	4

Scenes are grouped by arrangement, and every cycle sharing an arrangement goes to the same split. Four cycles of one arrangement spread across train and validation would leak, because they are four photographs of nearly the same scene.

The test split is sealed. Score it once, at the end, with the model and the prompt already fixed. Read it while tuning and it becomes a second validation set, and there is no third.

One disclosure: these same 96 frames were used to train a YOLO segmentation model in the originating project, and the test split here was drawn from that model's train side. For anyone training only on this dataset that makes no difference. It matters only if you run that particular detector into a model and score the pair here, which would read better than it should.

Two label variants, and why you want the strict one

Each target carries as_collected and strict_contact:

• strict_contact — two instruments are in contact when their measured outlines actually reach each other. Use this one. It is what answer.strict_contact_json is built from.
• as_collected — what the robot's planner recorded, which counts a pair as crossing anywhere within a ~14.45 mm margin. That margin is correct for deciding whether to release an instrument a few millimetres higher, and wrong as a claim about what is touching.

The difference is not small. Across the full capture, 44% of the pairs the planner recorded as crossing are not touching at all, and over half the scenes name fewer free instruments than the scene actually had. Train on as_collected and the model learns to assert, confidently, that two instruments visibly 13 mm apart are stacked.

Per-pair separations in millimetres are recorded under overlaps_provenance, so you can re-threshold without re-deriving anything.

Scoring: do not use exact-order accuracy

Several orders are correct for one tray — most of these scenes have three or more instruments that could be taken first — so a model naming an equally valid pick is marked wrong by exact match. It also has no gradient: an early run scored 0 of 9 and the number said nothing about whether it was improving.

Report instead:

• contact F1 (directed) — did it see which instrument lies under which
• contact F1 (undirected) — reported beside it, so finding a crossing and reversing it is separable from missing it
• order validity — is the order a legal clearing of the true stacking
• prefix validity — how many picks before it first digs underneath
• self-consistency — is the order legal under the contacts the model itself claimed; this separates misreading the scene from reasoning badly about what it saw

Baselines you have to beat

Measured on a held-out split, for answers that never looked at the image:

	claims nothing touches	invents a stack
contact F1 directed	0.25	0.07
order valid	0.69	0.44
prefix valid (of 5)	3.75	2.89
topmost valid	0.92	0.69
self-consistent	1.00	1.00

Read that topmost row before reporting one. A model that ignored the image entirely scores 0.92, because most trays have several valid first picks. Order validity is nearly as soft. Directed contact F1 is the number with real headroom, and the one to watch.

Limitations

• One camera, one table, one lighting regime. 25 arrangement groups from a single cell. Nothing here supports a claim about generalising to another setup — collect independent sessions before making one.
• Five instrument classes, one physical instance of each.
• One observation per scene, taken before any pick. There is no image of the tray after an instrument has been removed.
• Contact is modelled, not observed. It comes from measured instrument outlines placed at known poses, not from the camera. It is as good as those measurements.
• The stacking assumes a pick disturbs nothing else. The arm sets each instrument down from a small height, so this is not guaranteed, and nothing in the dataset can confirm or refute it.
• Masks can span an occluder. Where an instrument was split in two by something lying across it, the hand-corrected pieces were joined into a single ring. 52 such merges across 45 cycles, recorded per cycle.
• Small. 96 scenes is a proof of concept, not a benchmark.

Provenance

Each target carries overlaps_provenance: whether the contacts came from measured outlines or from a conservative fallback rule, with the clearance, per-instrument tolerances and angle-sign fit that decided it. All 96 came from measured outlines.

This is the hand-corrected subset of a larger capture of 181 cycles; the remainder carry detector masks of mixed quality and are not included here.

The motion log for each cycle is included; the video it aligns to is not.

Citation

@misc{iris_instrument_clearing_2026,
  title  = {IRIS surgical instrument clearing order},
  author = {Allelx},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Allelx/iris-instrument-clearing},
  note   = {96 robot-arranged scenes with logged stacking ground truth}
}