Shay85Gil/somethings_missing_here

Something's Missing Here

A dataset of viewpoint-aligned before/after image pairs for training and evaluating missing-object detection: each pair shows the same scene from (nearly) the same viewpoint, once with one or more objects present and once after they were removed, with a label naming what went missing.

The dataset originated from pairs derived from the simkoc/Remove360 dataset (training/Remove360_based/), extended with original photographs collected by the dataset author (training/DeTaken/), plus a held-out test/ split of scenes that appear nowhere in training. A separate challenging/ split holds harder cases — object swaps, replacements, and rearrangements — where telling "something is missing" apart from "things merely changed" takes more than spotting a difference (see below).

Structure

Both splits are organized by collection, then scene, then pair:

training/
├── DeTaken/                 original photographs by the dataset author
│   └── <scene>/             bullseye, cabinet_window, corner_table, michal_01,
│       └── pair_<nn>/       shay_01, vases, ...
│           ├── before.jpg   earlier image
│           ├── after.jpg    later image
│           └── label.json   {"missing": true, "items": ["<removed object>"]}
│                            or {"missing": false, "items": []} for no-change pairs
└── Remove360_based/         pairs derived from Remove360 (see below)
    └── <scene>/             backyard_big_tree, backyard_stones, park,
        └── pair_<nn>/       stairwell
test/
├── DeTaken/                 held-out scenes, never in training
│   └── test_<scene>/        test_class_01, test_class_02, test_class_03
│       └── pair_<nn>/
└── Remove360_based/         held-out Remove360 scenes, never in training
    └── test_<scene>/        test_backyard_bricks, test_backyard_toys,
        └── pair_<nn>/       test_bedroom, test_living-room, test_office
challenging/                 harder cases: swaps, replacements, rearrangements
├── training/
│   └── <scene>/             same scene names as training/DeTaken
│       └── pair_<nn>/
└── test/
    └── pair_<nn>/           held-out challenging pairs (flat)

Size in pairs: 350 training and 116 test (across 32 and 8 scenes respectively), plus 93 challenging pairs (81 training, 12 test) — 559 pairs in total. Breakdown (crop-shift = synthetic no-change pairs, see below):

split	collection	scenes	pairs	positive	negative	of which crop-shift
training	DeTaken	28	254	104	150	116
training	Remove360_based	4	96	48	48	48
test	DeTaken	3	40	21	19	0
test	Remove360_based	5	76	76	0	0
challenging	training	17	81	27	54	0
challenging	test	1	12	6	6	0

Test pairs are unmanipulated. Synthetic pairs are confined to training/: every pair in test/ and challenging/ consists of two genuinely captured photographs, with all removals physical — nothing was digitally added to, removed from, or composited into any image. Verified against the folder tree: no test label.json carries a source field (the marker of a derived pair), the crop-shift count is zero across all test splits, and pair folders contain nothing but before.jpg, after.jpg, and label.json. The only processing applied to test images is the disclosed geometric viewpoint alignment of the Remove360_based subset (a homography warp and crop — see below).

Negative pairs that were derived rather than photographed carry a source field in label.json documenting what they were built from.

Crop-shift negatives

Many scenes were photographed only with something removed, leaving no "nothing changed" pairs to learn from. Those scenes are supplemented with crop-shift negatives: both images of the pair are crops of the same photograph, taken at different offsets — one window trimmed at the right and top edges, the other trimmed by the same amounts at the bottom and left. The content is therefore identical and only the framing moves, which is exactly the "the camera shifted, the scene did not" case a removal detector must not mistake for a disappearance.

Two negatives are generated per source pair (one from its before.jpg, one from its after.jpg; the Remove360_based scenes retain only one of the two), the trim is 10–20 px per axis, both crops come out the same size, and each carries {"missing": false, "items": [], "source": "crop-shift of <pair>/<file>"}.

They are cheap and plentiful, but they only vary translation — unlike real re-shot no-change pairs, the lighting, focus, and perspective are identical. Treat them as a supplement: the crop-shift column above shows how much of each split's negative set is synthetic. Crop-shift pairs exist only in training/ — the test/ and challenging/ splits contain none, so a model that has merely learned "small shift ⇒ nothing missing" shows up there as false alarms.

In Remove360_based/ pairs (both splits), before.jpg is warped into after.jpg's camera frame and both images are cropped to their shared valid region, so the two images are pixel-aligned with identical dimensions. DeTaken/ and challenging/ pairs are handheld re-shots from approximately the same viewpoint and are not pixel-aligned. The class_* / test_class_* scenes are tabletop object arrangements shot in rapid succession (seconds apart), with roughly half of each scene being no-change negative pairs.

The challenging split

challenging/ holds the deliberately hard cases, separated from the main splits so models can be trained and evaluated with or without them. In the main splits, the change between before and after is a clean object removal (or nothing at all) seen under a small viewpoint change. In challenging/, other things happen too — objects are swapped with one another, replaced by different objects, or rearranged within the scene:

• Its negatives (missing: false, the majority here) are hard: items moved, swapped, or substituted — the scene visibly changed, yet nothing went missing. A model leaning on "the images differ, so something is gone" fails these.
• Its positives (missing: true) hide a genuine removal among such distractions, so spotting the difference is not enough — the model must identify that the change is specifically a disappearance.

Labels use the same label.json format. challenging/test/ pairs are held out from all training and sit directly under the folder (no scene subfolders).

The Remove360_based subset

Remove360 provides separate pre-removal and post-removal camera walks of real indoor and outdoor scenes. Its before and after images are independent captures — they are not pixel-aligned pairs — so this subset was built by finding and aligning the closest matching viewpoints between the two walks.

172 pairs — 124 positives (something was removed) and 48 crop-shift negatives — across 9 scenes and 9 removed objects, divided by whole scene into training/ (4 scenes, 96 pairs, one crop-shift negative per positive) and the held-out test/ (5 scenes, 76 pairs, positives only, prefixed test_). Remove360's single large backyard scene is split into four sub-scenes by area (big tree lawn, brick patio, stones, toy corner):

scene	positive	negative		removed object	pairs
stairwell	24	24		chairs	36
test_backyard_toys	23	0		backpack	24
test_living-room	21	0		stroller	16
test_backyard_bricks	16	0		sofa	15
test_office	13	0		deckchair	11
backyard_big_tree	11	11		bicycle	10
park	10	10		pillows	6
test_bedroom	3	0		table	3
backyard_stones	3	3		toy-truck	3

How it was generated

1. Download — the full simkoc/Remove360 repository (file tree of <scene>/<object>/train|test|masks), where train/ holds pre-removal ("before") images and test/ holds post-removal ("after") images.
2. Valid-region cropping — a subset of Remove360's images is truncated at fixed byte boundaries on the Hub itself (all of backyard/stroller at 2.75 MiB, all of backyard/playhouse at 256 KiB); truncated JPEGs decode with a uniform gray tail. Each image was cropped to its real content before matching, and images with less than 15% real content were discarded.
3. Viewpoint matching — every after image was ranked against all before images of the same object by SIFT feature matches (Lowe ratio 0.75); the top 3 candidates were verified with a RANSAC homography (reprojection threshold 4 px, minimum 40 inliers).
4. Acceptance criteria — a pair was kept only if each frame covers at least 85% of the other under the homography (mutual frame coverage) and the warped before image correlates with the after image at ≥ 0.475 zero-mean normalized correlation. Each before image was used in at most one pair.
5. Alignment and cropping — the accepted before image was warped into the after frame at full resolution and both images were cropped to the largest rectangle of shared valid pixels.
6. Labeling — each pair's label.json records the removed object (the Remove360 object folder the pair came from) as {"missing": true, "items": ["<object>"]}.
7. Manual curation — the automatically accepted pairs were reviewed and some were deleted by hand; the remaining pairs were renumbered contiguously.
8. No-change negatives — this subset's negatives are the crop-shift pairs described above, kept only in the training scenes (one per positive); the test scenes hold positives only. An earlier approach paired same-walk images (two pre-removal frames, or two post-removal frames, matched with the same gates as the positives) and a few such pairs may remain; their label.json source field names the two images they came from.

Known limitations

• No playhouse pairs — all of Remove360's backyard/playhouse images are truncated to ~6% of their content on the Hub, which is below the usability floor.
• Stroller pairs are half-height — backyard/stroller images are truncated to roughly their top half, so its pairs are wide bands (3900×1000) rather than full frames (4000×2200).
• Residual parallax — alignment uses a single homography per pair; small parallax between the two camera positions can remain, especially on close foreground geometry.
• Some objects other than the labeled one may have shifted slightly between Remove360's two capture sessions.

License and attribution

The Remove360_based/ subset is a derivative of simkoc/Remove360 and is distributed under the same CC-BY-NC-4.0 license, which this dataset adopts as a whole. If you use it, please also cite the original Remove360 paper (arXiv:2508.11431).