The measured selection test
The interiors above are showcase images. These results come from a separate set of 50 public photos used during development on 3 September 2026: 30 objects and 20 wall areas. The test measures selection cleanup, not finished renders or today’s product.
The September test version followed the labelled edges a little more closely on average. It also selected slightly more unwanted area. These two measurements need to be read together.
Edge match
How closely the selection follows the labelled outline. A score out of 100.
Original → September test: +2.34 points (better). 37 improved · 13 declined · 0 tied.
Extra area selected
The share of the selection that falls outside the labelled target.
Original → September test: +0.20 percentage points (worse). 20 improved · 30 declined · 0 tied.
Edge scores improved on 37 photos and fell on 13. Extra-area errors increased on 30 photos. An earlier cleanup version scored slightly better on all four averages than the September test version.
We did not adopt the proposed update. It changed none of the 50 selections, and the test still fell short of our recorded edge-match and missed-area requirements. The full results and thresholds are below.
Inspect three test photos and their selections
A small change, a larger improvement and a setback
These are examples from the measured set, chosen to show different outcomes. They are not a representative sample of all 50 photos.
Purple shows each saved selection; the purple-and-white line follows its boundary. Reference outline and Changed pixels are optional inspection views. The photos and mask geometry are unchanged. Each pair shows the full image at the same scale. Reconstructed after-selections match all four recorded scores exactly.
A small change around the oven
Edge match 82.37 /100Extra area selected 4.34%
Edge match 83.96 /100Extra area selected 4.21%
Changed pixels: 169. Pink marks pixels added to or removed from the selection, at their exact size. Both photos show the same diagnostic. Other overlays are hidden until this view is switched off.
The target is the oven on the left. Cleanup changes which 169 pixels are selected, with a modest increase in the edge score. Most of the selection stays the same.
Why this example: A change close to the middle of the 50 recorded edge-score changes.
Photo: Kitchen YHA Rowen by Eric Audige-Soutter, CC BY 2.0. Unmodified COCO source photo; selection overlays added by Rendero. Reference outline rendered as SVG from COCO Consortium labels, CC BY 4.0.
Photo and selection record
COCO image 175364 · 614 × 461 pixels · 169 selection pixels changed. The after-selection is a deterministic reconstruction from the saved inputs and historical cleanup code; every remeasured score matches the original record exactly.
Open the original photo · Original selection · Tested selection · Dataset reference · Full photo evidence record
Fewer gaps in the wall selection
Edge match 26.21 /100Extra area selected 2.74%
Edge match 55.23 /100Extra area selected 2.35%
Changed pixels: 9,492. Pink marks pixels added to or removed from the selection, at their exact size. Both photos show the same diagnostic. Other overlays are hidden until this view is switched off.
Look at the wall behind the boat. Cleanup fills many small gaps in the selected area. Its edge score improves, although part of the reference area is still missed.
Why this example: The largest edge-score improvement in this test; it is an example of a strong result.
Photo: Parked Boat by Charles Hutchins, CC BY 2.0. Unmodified COCO source photo; selection overlays added by Rendero. Reference outline rendered as SVG from Holger Caesar, Jasper Uijlings and Vittorio Ferrari / COCO-Stuff labels, CC BY 4.0.
Photo and selection record
COCO image 113051 · 640 × 480 pixels · 9,492 selection pixels changed. The after-selection is a deterministic reconstruction from the saved inputs and historical cleanup code; every remeasured score matches the original record exactly.
Open the original photo · Original selection · Tested selection · Dataset reference · Full photo evidence record
An edge score that went backwards
Edge match 69.10 /100Extra area selected 3.30%
Edge match 64.26 /100Extra area selected 3.44%
Changed pixels: 221. Pink marks pixels added to or removed from the selection, at their exact size. Both photos show the same diagnostic. Other overlays are hidden until this view is switched off.
The refrigerator is on the far left. Both selections look similar at this scale, but 221 pixels change selection status and the edge score falls. The numbers catch a setback that is easy to miss at a glance.
Why this example: A case with a lower edge score after cleanup, included to show a setback.
Photo: Kitchen by marysalome, CC BY 2.0. Unmodified COCO source photo; selection overlays added by Rendero. Reference outline rendered as SVG from COCO Consortium labels, CC BY 4.0.
Photo and selection record
COCO image 429598 · 640 × 480 pixels · 221 selection pixels changed. The after-selection is a deterministic reconstruction from the saved inputs and historical cleanup code; every remeasured score matches the original record exactly.
Open the original photo · Original selection · Tested selection · Dataset reference · Full photo evidence record
Compare all five versions and four measurements
All charts use a fixed zero-to-100 scale. Scores are averages with equal weight for each photo. Filter objects and walls to inspect a subset.
Showing all 50 cases. Charts and per-case rows use the same selection.
Area match
How closely the selection overlaps the labelled target.
Original → September test: +0.39 percentage points (better). 35 improved · 15 declined · 0 tied.
Edge match
How closely the selection follows the labelled outline. A score out of 100.
Original → September test: +2.34 points (better). 37 improved · 13 declined · 0 tied.
Extra area selected
The share of the selection that falls outside the labelled target.
Original → September test: +0.20 percentage points (worse). 20 improved · 30 declined · 0 tied.
Target area missed
The share of the labelled target left out of the selection.
Original → September test: −0.42 percentage points (better). 42 improved · 8 declined · 0 tied.
The shape-check version, September test version and proposed update have identical scores on this set. Version identifiers are recorded in the technical details below.
How we tested
- Test date
- 3 September 2026. We replayed cleanup on saved selections; the image model was not run again.
- Sample
- 50 distinct COCO validation images: 30 object instances and 20 COCO-Stuff wall surfaces. All reused in development.
- Same starting point
- Every cleanup version received the same original photo, saved selection and model confidence value.
- Image resolution
- Native widths 320–640 px and heights 240–640 px. No resizing during replay.
- What counts as correct
- The labelled target in COCO or COCO-Stuff. Its location was supplied in advance, with 5% padding around its box. This tests cleanup after an object has been located.
- Limits
- These photos had already informed development. They do not establish performance on new projects. We did not measure finished renders, automatic object finding, speed, cost or other products.
Technical details: scores, versions and test requirements
The saved image model was SAM 2 Hiera Large via RSAM, with mode sam2_compat and operation parent_masks_v1. These identifiers describe the historical test configuration.
Chart names and source versions: Original selection = saved raw parent mask; Earlier cleanup = 9317364; Cleanup with shape checks = 91b8f00; September test version = 5ceddcf; Proposed update = the unpromoted line-recovery candidate. Exact source hashes and encoder/decoder revisions are in the full specification.
Test requirements: the proposed update’s mean edge score was 51.47 /100, below the required 68.00 /100. Its missed-area rate was 20.47%, above the 18.00% limit. These are full-set values; the complete registered requirements are in the specification.
Reader-facing names map to the recorded metrics as follows: area match = intersection over union (IoU); edge match = boundary F1; extra area selected = leakage; target area missed = missed rate.
Prediction masks are binarized above 127. Reference masks have the same native dimensions. Source scores range from 0 to 1. Edge match is shown out of 100; other ratios are shown as percentages.
IoU = intersection pixels / union pixels
Leakage = false-positive pixels / predicted foreground pixels
Missed area = false-negative pixels / reference foreground pixels
Each denominator is clamped to at least 1. Boundary F1 is the harmonic mean of boundary precision and recall. Boundaries use four-connected erosion; the outer image border counts as boundary. Matches use a 5 × 5 square max filter: Chebyshev distance ≤ 2 native pixels. Two empty boundaries score 1; otherwise an absent boundary receives zero match credit.
Reported mean F1 is the average of the 50 individual F1 scores. It is not F1 calculated from pooled or averaged precision and recall. Paired changes within 10−12 count as ties. Changes on the out-of-100 edge scale are points; changes in percentage ratios are percentage points. Both equal the difference in the original scores multiplied by 100.
Versions, input provenance and reproducibility limits
The historical refiners are identified by revisions 9317364, 91b8f00 and 5ceddcf. The opt-in recovery candidate was not deployed. Exact refiner-source hashes and the saved upstream encoder and decoder revisions are recorded in the full specification.
Reference pixel masks were not passed to inference, but their bounds supplied the localization boxes. This controls localization out of the comparison. COCO labels are independent of the predictions, but they do not provide detailed CAD geometry or fine alpha-matte truth.
The fixture originally assigned 30 diagnosis cases and 20 holdouts. All 50 were later reused; those labels now describe provenance only. Results are descriptive, with no untouched confirmation set, confidence interval or statistical-significance claim.
All 50 saved-response hashes were checked. The exported per-case records independently reproduce all 20 recorded mean scores and 12 paired comparisons. For the three photo examples only, we also reconstructed the September cleanup from the saved inputs and original source code, then remeasured all four scores: each matches the saved record exactly. This did not rerun image-model inference or independently confirm the remaining 47 cases from pixels.
Hardware, the complete model-weight digest and full upstream inference settings were not recorded. The score downloads contain derived results and metadata; the three licensed photo examples and separate SVG overlays are additional evidence on this page.
Every case, including the regressions
The table includes all five stages for every image. Filters above apply to these rows too. Full-precision values and reference identifiers are available in the downloads.
Open the per-case score table · 250 rows for 50 images
| COCO image | Category | Track | Version | Area match ↑ | Edge match /100 ↑ | Extra area ↓ | Missed area ↓ |
|---|---|---|---|---|---|---|---|
| 1993 | bed | object | Original selection | 86.66% | 51.36 /100 | 0.46% | 12.99% |
| 1993 | bed | object | Earlier cleanup | 86.54% | 52.39 /100 | 0.48% | 13.10% |
| 1993 | bed | object | Cleanup with shape checks | 86.54% | 52.39 /100 | 0.48% | 13.10% |
| 1993 | bed | object | September test version | 86.54% | 52.39 /100 | 0.48% | 13.10% |
| 1993 | bed | object | Proposed update · not adopted | 86.54% | 52.39 /100 | 0.48% | 13.10% |
| 10092 | bed | object | Original selection | 5.59% | 21.30 /100 | 7.95% | 94.39% |
| 10092 | bed | object | Earlier cleanup | 5.64% | 22.10 /100 | 7.72% | 94.33% |
| 10092 | bed | object | Cleanup with shape checks | 5.64% | 22.10 /100 | 7.72% | 94.33% |
| 10092 | bed | object | September test version | 5.64% | 22.10 /100 | 7.72% | 94.33% |
| 10092 | bed | object | Proposed update · not adopted | 5.64% | 22.10 /100 | 7.72% | 94.33% |
| 464358 | bed | object | Original selection | 63.85% | 9.04 /100 | 11.43% | 30.41% |
| 464358 | bed | object | Earlier cleanup | 64.02% | 8.62 /100 | 11.40% | 30.24% |
| 464358 | bed | object | Cleanup with shape checks | 64.02% | 8.62 /100 | 11.40% | 30.24% |
| 464358 | bed | object | September test version | 64.02% | 8.62 /100 | 11.40% | 30.24% |
| 464358 | bed | object | Proposed update · not adopted | 64.02% | 8.62 /100 | 11.40% | 30.24% |
| 340451 | bench | object | Original selection | 60.06% | 55.90 /100 | 1.52% | 39.38% |
| 340451 | bench | object | Earlier cleanup | 60.48% | 57.96 /100 | 1.48% | 38.97% |
| 340451 | bench | object | Cleanup with shape checks | 60.20% | 57.28 /100 | 1.49% | 39.25% |
| 340451 | bench | object | September test version | 60.20% | 57.28 /100 | 1.49% | 39.25% |
| 340451 | bench | object | Proposed update · not adopted | 60.20% | 57.28 /100 | 1.49% | 39.25% |
| 221502 | bench | object | Original selection | 82.86% | 63.20 /100 | 0.18% | 17.02% |
| 221502 | bench | object | Earlier cleanup | 83.32% | 64.43 /100 | 0.22% | 16.53% |
| 221502 | bench | object | Cleanup with shape checks | 83.32% | 64.43 /100 | 0.22% | 16.53% |
| 221502 | bench | object | September test version | 83.32% | 64.43 /100 | 0.22% | 16.53% |
| 221502 | bench | object | Proposed update · not adopted | 83.32% | 64.43 /100 | 0.22% | 16.53% |
| 420916 | bench | object | Original selection | 48.86% | 46.51 /100 | 4.57% | 49.97% |
| 420916 | bench | object | Earlier cleanup | 50.14% | 49.71 /100 | 4.10% | 48.76% |
| 420916 | bench | object | Cleanup with shape checks | 49.26% | 48.37 /100 | 4.17% | 49.66% |
| 420916 | bench | object | September test version | 49.26% | 48.37 /100 | 4.17% | 49.66% |
| 420916 | bench | object | Proposed update · not adopted | 49.26% | 48.37 /100 | 4.17% | 49.66% |
| 27620 | chair | object | Original selection | 90.41% | 62.55 /100 | 5.80% | 4.27% |
| 27620 | chair | object | Earlier cleanup | 90.06% | 60.48 /100 | 6.21% | 4.23% |
| 27620 | chair | object | Cleanup with shape checks | 90.06% | 60.48 /100 | 6.21% | 4.23% |
| 27620 | chair | object | September test version | 90.06% | 60.48 /100 | 6.21% | 4.23% |
| 27620 | chair | object | Proposed update · not adopted | 90.06% | 60.48 /100 | 6.21% | 4.23% |
| 546964 | chair | object | Original selection | 92.15% | 51.36 /100 | 3.32% | 4.84% |
| 546964 | chair | object | Earlier cleanup | 92.71% | 52.96 /100 | 3.37% | 4.19% |
| 546964 | chair | object | Cleanup with shape checks | 92.08% | 51.20 /100 | 3.39% | 4.85% |
| 546964 | chair | object | September test version | 92.08% | 51.20 /100 | 3.39% | 4.85% |
| 546964 | chair | object | Proposed update · not adopted | 92.08% | 51.20 /100 | 3.39% | 4.85% |
| 189820 | chair | object | Original selection | 73.59% | 36.61 /100 | 3.53% | 24.38% |
| 189820 | chair | object | Earlier cleanup | 73.43% | 37.10 /100 | 3.33% | 24.67% |
| 189820 | chair | object | Cleanup with shape checks | 73.43% | 37.10 /100 | 3.33% | 24.67% |
| 189820 | chair | object | September test version | 73.43% | 37.10 /100 | 3.33% | 24.67% |
| 189820 | chair | object | Proposed update · not adopted | 73.43% | 37.10 /100 | 3.33% | 24.67% |
| 356432 | couch | object | Original selection | 79.73% | 56.45 /100 | 0.89% | 19.69% |
| 356432 | couch | object | Earlier cleanup | 80.01% | 58.38 /100 | 0.94% | 19.38% |
| 356432 | couch | object | Cleanup with shape checks | 80.01% | 58.38 /100 | 0.94% | 19.38% |
| 356432 | couch | object | September test version | 80.01% | 58.38 /100 | 0.94% | 19.38% |
| 356432 | couch | object | Proposed update · not adopted | 80.01% | 58.38 /100 | 0.94% | 19.38% |
| 31735 | couch | object | Original selection | 89.80% | 61.10 /100 | 0.97% | 9.40% |
| 31735 | couch | object | Earlier cleanup | 90.25% | 63.52 /100 | 1.10% | 8.84% |
| 31735 | couch | object | Cleanup with shape checks | 90.25% | 63.52 /100 | 1.10% | 8.84% |
| 31735 | couch | object | September test version | 90.25% | 63.52 /100 | 1.10% | 8.84% |
| 31735 | couch | object | Proposed update · not adopted | 90.25% | 63.52 /100 | 1.10% | 8.84% |
| 232088 | couch | object | Original selection | 84.18% | 17.44 /100 | 4.71% | 12.17% |
| 232088 | couch | object | Earlier cleanup | 86.16% | 19.02 /100 | 4.75% | 9.97% |
| 232088 | couch | object | Cleanup with shape checks | 84.06% | 17.14 /100 | 4.87% | 12.16% |
| 232088 | couch | object | September test version | 84.06% | 17.14 /100 | 4.87% | 12.16% |
| 232088 | couch | object | Proposed update · not adopted | 84.06% | 17.14 /100 | 4.87% | 12.16% |
| 398652 | dining table | object | Original selection | 72.26% | 35.29 /100 | 12.05% | 19.80% |
| 398652 | dining table | object | Earlier cleanup | 73.17% | 43.26 /100 | 10.83% | 19.69% |
| 398652 | dining table | object | Cleanup with shape checks | 73.17% | 43.26 /100 | 10.83% | 19.69% |
| 398652 | dining table | object | September test version | 73.17% | 43.26 /100 | 10.83% | 19.69% |
| 398652 | dining table | object | Proposed update · not adopted | 73.17% | 43.26 /100 | 10.83% | 19.69% |
| 397133 | dining table | object | Original selection | 50.72% | 30.51 /100 | 2.69% | 48.56% |
| 397133 | dining table | object | Earlier cleanup | 50.84% | 31.32 /100 | 2.63% | 48.45% |
| 397133 | dining table | object | Cleanup with shape checks | 50.84% | 31.32 /100 | 2.63% | 48.45% |
| 397133 | dining table | object | September test version | 50.84% | 31.32 /100 | 2.63% | 48.45% |
| 397133 | dining table | object | Proposed update · not adopted | 50.84% | 31.32 /100 | 2.63% | 48.45% |
| 256916 | dining table | object | Original selection | 87.71% | 70.35 /100 | 7.45% | 5.62% |
| 256916 | dining table | object | Earlier cleanup | 87.86% | 71.92 /100 | 8.04% | 4.83% |
| 256916 | dining table | object | Cleanup with shape checks | 87.86% | 71.92 /100 | 8.04% | 4.83% |
| 256916 | dining table | object | September test version | 87.86% | 71.92 /100 | 8.04% | 4.83% |
| 256916 | dining table | object | Proposed update · not adopted | 87.86% | 71.92 /100 | 8.04% | 4.83% |
| 365766 | oven | object | Original selection | 85.43% | 63.71 /100 | 10.50% | 5.06% |
| 365766 | oven | object | Earlier cleanup | 85.37% | 63.63 /100 | 10.58% | 5.03% |
| 365766 | oven | object | Cleanup with shape checks | 85.37% | 63.63 /100 | 10.58% | 5.03% |
| 365766 | oven | object | September test version | 85.37% | 63.63 /100 | 10.58% | 5.03% |
| 365766 | oven | object | Proposed update · not adopted | 85.37% | 63.63 /100 | 10.58% | 5.03% |
| 175364 | oven | object | Original selection | 94.09% | 82.37 /100 | 4.34% | 1.72% |
| 175364 | oven | object | Earlier cleanup | 94.37% | 83.96 /100 | 4.21% | 1.55% |
| 175364 | oven | object | Cleanup with shape checks | 94.37% | 83.96 /100 | 4.21% | 1.55% |
| 175364 | oven | object | September test version | 94.37% | 83.96 /100 | 4.21% | 1.55% |
| 175364 | oven | object | Proposed update · not adopted | 94.37% | 83.96 /100 | 4.21% | 1.55% |
| 424349 | oven | object | Original selection | 82.05% | 49.77 /100 | 3.39% | 15.51% |
| 424349 | oven | object | Earlier cleanup | 82.54% | 58.69 /100 | 3.35% | 15.03% |
| 424349 | oven | object | Cleanup with shape checks | 82.54% | 58.69 /100 | 3.35% | 15.03% |
| 424349 | oven | object | September test version | 82.54% | 58.69 /100 | 3.35% | 15.03% |
| 424349 | oven | object | Proposed update · not adopted | 82.54% | 58.69 /100 | 3.35% | 15.03% |
| 437514 | potted plant | object | Original selection | 76.02% | 12.63 /100 | 3.24% | 22.00% |
| 437514 | potted plant | object | Earlier cleanup | 76.74% | 14.28 /100 | 3.08% | 21.35% |
| 437514 | potted plant | object | Cleanup with shape checks | 76.52% | 14.00 /100 | 3.08% | 21.57% |
| 437514 | potted plant | object | September test version | 76.52% | 14.00 /100 | 3.08% | 21.57% |
| 437514 | potted plant | object | Proposed update · not adopted | 76.52% | 14.00 /100 | 3.08% | 21.57% |
| 569700 | potted plant | object | Original selection | 80.15% | 31.55 /100 | 8.75% | 13.19% |
| 569700 | potted plant | object | Earlier cleanup | 80.77% | 36.65 /100 | 8.65% | 12.54% |
| 569700 | potted plant | object | Cleanup with shape checks | 80.10% | 35.40 /100 | 8.71% | 13.27% |
| 569700 | potted plant | object | September test version | 80.10% | 35.40 /100 | 8.71% | 13.27% |
| 569700 | potted plant | object | Proposed update · not adopted | 80.10% | 35.40 /100 | 8.71% | 13.27% |
| 404484 | potted plant | object | Original selection | 0.27% | 13.10 /100 | 95.89% | 99.71% |
| 404484 | potted plant | object | Earlier cleanup | 0.17% | 5.42 /100 | 97.21% | 99.82% |
| 404484 | potted plant | object | Cleanup with shape checks | 0.17% | 5.42 /100 | 97.21% | 99.82% |
| 404484 | potted plant | object | September test version | 0.17% | 5.42 /100 | 97.21% | 99.82% |
| 404484 | potted plant | object | Proposed update · not adopted | 0.17% | 5.42 /100 | 97.21% | 99.82% |
| 429598 | refrigerator | object | Original selection | 96.11% | 69.10 /100 | 3.30% | 0.63% |
| 429598 | refrigerator | object | Earlier cleanup | 96.07% | 64.26 /100 | 3.44% | 0.53% |
| 429598 | refrigerator | object | Cleanup with shape checks | 96.07% | 64.26 /100 | 3.44% | 0.53% |
| 429598 | refrigerator | object | September test version | 96.07% | 64.26 /100 | 3.44% | 0.53% |
| 429598 | refrigerator | object | Proposed update · not adopted | 96.07% | 64.26 /100 | 3.44% | 0.53% |
| 313130 | refrigerator | object | Original selection | 96.00% | 65.78 /100 | 3.10% | 0.95% |
| 313130 | refrigerator | object | Earlier cleanup | 96.18% | 68.06 /100 | 3.10% | 0.77% |
| 313130 | refrigerator | object | Cleanup with shape checks | 96.18% | 68.06 /100 | 3.10% | 0.77% |
| 313130 | refrigerator | object | September test version | 96.18% | 68.06 /100 | 3.10% | 0.77% |
| 313130 | refrigerator | object | Proposed update · not adopted | 96.18% | 68.06 /100 | 3.10% | 0.77% |
| 52996 | refrigerator | object | Original selection | 95.30% | 76.47 /100 | 1.32% | 3.47% |
| 52996 | refrigerator | object | Earlier cleanup | 95.20% | 72.96 /100 | 1.52% | 3.38% |
| 52996 | refrigerator | object | Cleanup with shape checks | 95.20% | 72.96 /100 | 1.52% | 3.38% |
| 52996 | refrigerator | object | September test version | 95.20% | 72.96 /100 | 1.52% | 3.38% |
| 52996 | refrigerator | object | Proposed update · not adopted | 95.20% | 72.96 /100 | 1.52% | 3.38% |
| 241319 | sink | object | Original selection | 92.10% | 68.99 /100 | 5.76% | 2.41% |
| 241319 | sink | object | Earlier cleanup | 91.90% | 71.08 /100 | 6.13% | 2.23% |
| 241319 | sink | object | Cleanup with shape checks | 91.90% | 71.08 /100 | 6.13% | 2.23% |
| 241319 | sink | object | September test version | 91.90% | 71.08 /100 | 6.13% | 2.23% |
| 241319 | sink | object | Proposed update · not adopted | 91.90% | 71.08 /100 | 6.13% | 2.23% |
| 569917 | sink | object | Original selection | 66.08% | 13.55 /100 | 19.58% | 21.24% |
| 569917 | sink | object | Earlier cleanup | 67.25% | 12.74 /100 | 19.48% | 19.68% |
| 569917 | sink | object | Cleanup with shape checks | 66.64% | 12.35 /100 | 19.62% | 20.40% |
| 569917 | sink | object | September test version | 66.64% | 12.35 /100 | 19.62% | 20.40% |
| 569917 | sink | object | Proposed update · not adopted | 66.64% | 12.35 /100 | 19.62% | 20.40% |
| 223789 | sink | object | Original selection | 90.09% | 39.18 /100 | 7.85% | 2.42% |
| 223789 | sink | object | Earlier cleanup | 90.03% | 43.45 /100 | 7.93% | 2.41% |
| 223789 | sink | object | Cleanup with shape checks | 90.03% | 43.45 /100 | 7.93% | 2.41% |
| 223789 | sink | object | September test version | 90.03% | 43.45 /100 | 7.93% | 2.41% |
| 223789 | sink | object | Proposed update · not adopted | 90.03% | 43.45 /100 | 7.93% | 2.41% |
| 371529 | toilet | object | Original selection | 88.62% | 51.97 /100 | 3.05% | 8.84% |
| 371529 | toilet | object | Earlier cleanup | 88.93% | 56.07 /100 | 2.86% | 8.68% |
| 371529 | toilet | object | Cleanup with shape checks | 88.93% | 56.07 /100 | 2.86% | 8.68% |
| 371529 | toilet | object | September test version | 88.93% | 56.07 /100 | 2.86% | 8.68% |
| 371529 | toilet | object | Proposed update · not adopted | 88.93% | 56.07 /100 | 2.86% | 8.68% |
| 234413 | toilet | object | Original selection | 93.91% | 59.82 /100 | 1.78% | 4.46% |
| 234413 | toilet | object | Earlier cleanup | 94.04% | 57.16 /100 | 1.85% | 4.27% |
| 234413 | toilet | object | Cleanup with shape checks | 94.04% | 57.16 /100 | 1.85% | 4.27% |
| 234413 | toilet | object | September test version | 94.04% | 57.16 /100 | 1.85% | 4.27% |
| 234413 | toilet | object | Proposed update · not adopted | 94.04% | 57.16 /100 | 1.85% | 4.27% |
| 537812 | toilet | object | Original selection | 95.21% | 87.91 /100 | 3.14% | 1.76% |
| 537812 | toilet | object | Earlier cleanup | 95.07% | 87.65 /100 | 2.94% | 2.11% |
| 537812 | toilet | object | Cleanup with shape checks | 95.07% | 87.65 /100 | 2.94% | 2.11% |
| 537812 | toilet | object | September test version | 95.07% | 87.65 /100 | 2.94% | 2.11% |
| 537812 | toilet | object | Proposed update · not adopted | 95.07% | 87.65 /100 | 2.94% | 2.11% |
| 39956 | wall-brick | surface | Original selection | 95.48% | 66.39 /100 | 1.05% | 3.54% |
| 39956 | wall-brick | surface | Earlier cleanup | 95.72% | 66.98 /100 | 1.09% | 3.26% |
| 39956 | wall-brick | surface | Cleanup with shape checks | 95.72% | 66.98 /100 | 1.09% | 3.26% |
| 39956 | wall-brick | surface | September test version | 95.72% | 66.98 /100 | 1.09% | 3.26% |
| 39956 | wall-brick | surface | Proposed update · not adopted | 95.72% | 66.98 /100 | 1.09% | 3.26% |
| 124659 | wall-brick | surface | Original selection | 66.10% | 45.41 /100 | 21.22% | 19.58% |
| 124659 | wall-brick | surface | Earlier cleanup | 66.96% | 51.57 /100 | 21.62% | 17.87% |
| 124659 | wall-brick | surface | Cleanup with shape checks | 66.96% | 51.57 /100 | 21.62% | 17.87% |
| 124659 | wall-brick | surface | September test version | 66.96% | 51.57 /100 | 21.62% | 17.87% |
| 124659 | wall-brick | surface | Proposed update · not adopted | 66.96% | 51.57 /100 | 21.62% | 17.87% |
| 281032 | wall-brick | surface | Original selection | 93.10% | 65.83 /100 | 0.46% | 6.49% |
| 281032 | wall-brick | surface | Earlier cleanup | 93.27% | 65.93 /100 | 0.54% | 6.25% |
| 281032 | wall-brick | surface | Cleanup with shape checks | 93.27% | 65.93 /100 | 0.54% | 6.25% |
| 281032 | wall-brick | surface | September test version | 93.27% | 65.93 /100 | 0.54% | 6.25% |
| 281032 | wall-brick | surface | Proposed update · not adopted | 93.27% | 65.93 /100 | 0.54% | 6.25% |
| 415194 | wall-concrete | surface | Original selection | 88.04% | 58.20 /100 | 3.81% | 8.78% |
| 415194 | wall-concrete | surface | Earlier cleanup | 89.07% | 60.34 /100 | 3.91% | 7.58% |
| 415194 | wall-concrete | surface | Cleanup with shape checks | 89.07% | 60.34 /100 | 3.91% | 7.58% |
| 415194 | wall-concrete | surface | September test version | 89.07% | 60.34 /100 | 3.91% | 7.58% |
| 415194 | wall-concrete | surface | Proposed update · not adopted | 89.07% | 60.34 /100 | 3.91% | 7.58% |
| 148620 | wall-concrete | surface | Original selection | 96.67% | 87.51 /100 | 0.52% | 2.84% |
| 148620 | wall-concrete | surface | Earlier cleanup | 97.04% | 89.24 /100 | 0.59% | 2.40% |
| 148620 | wall-concrete | surface | Cleanup with shape checks | 97.04% | 89.24 /100 | 0.59% | 2.40% |
| 148620 | wall-concrete | surface | September test version | 97.04% | 89.24 /100 | 0.59% | 2.40% |
| 148620 | wall-concrete | surface | Proposed update · not adopted | 97.04% | 89.24 /100 | 0.59% | 2.40% |
| 367569 | wall-concrete | surface | Original selection | 78.68% | 46.81 /100 | 0.93% | 20.74% |
| 367569 | wall-concrete | surface | Earlier cleanup | 79.27% | 48.24 /100 | 1.03% | 20.06% |
| 367569 | wall-concrete | surface | Cleanup with shape checks | 79.27% | 48.24 /100 | 1.03% | 20.06% |
| 367569 | wall-concrete | surface | September test version | 79.27% | 48.24 /100 | 1.03% | 20.06% |
| 367569 | wall-concrete | surface | Proposed update · not adopted | 79.27% | 48.24 /100 | 1.03% | 20.06% |
| 292005 | wall-other | surface | Original selection | 90.39% | 58.40 /100 | 7.71% | 2.23% |
| 292005 | wall-other | surface | Earlier cleanup | 90.58% | 64.51 /100 | 7.89% | 1.80% |
| 292005 | wall-other | surface | Cleanup with shape checks | 90.58% | 64.51 /100 | 7.89% | 1.80% |
| 292005 | wall-other | surface | September test version | 90.58% | 64.51 /100 | 7.89% | 1.80% |
| 292005 | wall-other | surface | Proposed update · not adopted | 90.58% | 64.51 /100 | 7.89% | 1.80% |
| 329080 | wall-other | surface | Original selection | 94.85% | 81.04 /100 | 0.85% | 4.37% |
| 329080 | wall-other | surface | Earlier cleanup | 95.00% | 80.92 /100 | 1.02% | 4.06% |
| 329080 | wall-other | surface | Cleanup with shape checks | 95.00% | 80.92 /100 | 1.02% | 4.06% |
| 329080 | wall-other | surface | September test version | 95.00% | 80.92 /100 | 1.02% | 4.06% |
| 329080 | wall-other | surface | Proposed update · not adopted | 95.00% | 80.92 /100 | 1.02% | 4.06% |
| 577584 | wall-other | surface | Original selection | 94.18% | 80.79 /100 | 1.96% | 4.02% |
| 577584 | wall-other | surface | Earlier cleanup | 94.46% | 82.38 /100 | 2.09% | 3.60% |
| 577584 | wall-other | surface | Cleanup with shape checks | 94.46% | 82.38 /100 | 2.09% | 3.60% |
| 577584 | wall-other | surface | September test version | 94.46% | 82.38 /100 | 2.09% | 3.60% |
| 577584 | wall-other | surface | Proposed update · not adopted | 94.46% | 82.38 /100 | 2.09% | 3.60% |
| 529762 | wall-panel | surface | Original selection | 45.74% | 37.38 /100 | 0.49% | 54.16% |
| 529762 | wall-panel | surface | Earlier cleanup | 46.21% | 44.31 /100 | 0.25% | 53.74% |
| 529762 | wall-panel | surface | Cleanup with shape checks | 46.21% | 44.31 /100 | 0.25% | 53.74% |
| 529762 | wall-panel | surface | September test version | 46.21% | 44.31 /100 | 0.25% | 53.74% |
| 529762 | wall-panel | surface | Proposed update · not adopted | 46.21% | 44.31 /100 | 0.25% | 53.74% |
| 13659 | wall-panel | surface | Original selection | 48.62% | 29.30 /100 | 15.16% | 46.76% |
| 13659 | wall-panel | surface | Earlier cleanup | 50.48% | 32.52 /100 | 11.96% | 45.80% |
| 13659 | wall-panel | surface | Cleanup with shape checks | 50.48% | 32.52 /100 | 11.96% | 45.80% |
| 13659 | wall-panel | surface | September test version | 50.48% | 32.52 /100 | 11.96% | 45.80% |
| 13659 | wall-panel | surface | Proposed update · not adopted | 50.48% | 32.52 /100 | 11.96% | 45.80% |
| 75612 | wall-panel | surface | Original selection | 94.82% | 70.32 /100 | 3.81% | 1.47% |
| 75612 | wall-panel | surface | Earlier cleanup | 95.07% | 71.92 /100 | 3.75% | 1.27% |
| 75612 | wall-panel | surface | Cleanup with shape checks | 95.07% | 71.92 /100 | 3.75% | 1.27% |
| 75612 | wall-panel | surface | September test version | 95.07% | 71.92 /100 | 3.75% | 1.27% |
| 75612 | wall-panel | surface | Proposed update · not adopted | 95.07% | 71.92 /100 | 3.75% | 1.27% |
| 65350 | wall-stone | surface | Original selection | 80.43% | 37.13 /100 | 11.83% | 9.83% |
| 65350 | wall-stone | surface | Earlier cleanup | 81.16% | 45.37 /100 | 11.84% | 8.91% |
| 65350 | wall-stone | surface | Cleanup with shape checks | 81.16% | 45.37 /100 | 11.84% | 8.91% |
| 65350 | wall-stone | surface | September test version | 81.16% | 45.37 /100 | 11.84% | 8.91% |
| 65350 | wall-stone | surface | Proposed update · not adopted | 81.16% | 45.37 /100 | 11.84% | 8.91% |
| 113051 | wall-stone | surface | Original selection | 78.91% | 26.21 /100 | 2.74% | 19.29% |
| 113051 | wall-stone | surface | Earlier cleanup | 84.69% | 63.63 /100 | 2.31% | 13.58% |
| 113051 | wall-stone | surface | Cleanup with shape checks | 83.27% | 55.23 /100 | 2.35% | 15.03% |
| 113051 | wall-stone | surface | September test version | 83.27% | 55.23 /100 | 2.35% | 15.03% |
| 113051 | wall-stone | surface | Proposed update · not adopted | 83.27% | 55.23 /100 | 2.35% | 15.03% |
| 526103 | wall-stone | surface | Original selection | 85.79% | 52.67 /100 | 1.54% | 13.04% |
| 526103 | wall-stone | surface | Earlier cleanup | 86.68% | 59.60 /100 | 2.02% | 11.74% |
| 526103 | wall-stone | surface | Cleanup with shape checks | 86.40% | 58.58 /100 | 2.02% | 12.03% |
| 526103 | wall-stone | surface | September test version | 86.40% | 58.58 /100 | 2.02% | 12.03% |
| 526103 | wall-stone | surface | Proposed update · not adopted | 86.40% | 58.58 /100 | 2.02% | 12.03% |
| 393569 | wall-tile | surface | Original selection | 74.74% | 42.55 /100 | 19.48% | 8.75% |
| 393569 | wall-tile | surface | Earlier cleanup | 76.53% | 46.55 /100 | 18.51% | 7.37% |
| 393569 | wall-tile | surface | Cleanup with shape checks | 76.53% | 46.55 /100 | 18.51% | 7.37% |
| 393569 | wall-tile | surface | September test version | 76.53% | 46.55 /100 | 18.51% | 7.37% |
| 393569 | wall-tile | surface | Proposed update · not adopted | 76.53% | 46.55 /100 | 18.51% | 7.37% |
| 570736 | wall-tile | surface | Original selection | 68.62% | 35.11 /100 | 18.30% | 18.92% |
| 570736 | wall-tile | surface | Earlier cleanup | 71.09% | 42.60 /100 | 17.32% | 16.47% |
| 570736 | wall-tile | surface | Cleanup with shape checks | 70.16% | 41.70 /100 | 17.51% | 17.56% |
| 570736 | wall-tile | surface | September test version | 70.16% | 41.70 /100 | 17.51% | 17.56% |
| 570736 | wall-tile | surface | Proposed update · not adopted | 70.16% | 41.70 /100 | 17.51% | 17.56% |
| 90108 | wall-tile | surface | Original selection | 0.07% | 9.10 /100 | 99.79% | 99.90% |
| 90108 | wall-tile | surface | Earlier cleanup | 0.02% | 11.20 /100 | 99.94% | 99.97% |
| 90108 | wall-tile | surface | Cleanup with shape checks | 0.02% | 11.20 /100 | 99.94% | 99.97% |
| 90108 | wall-tile | surface | September test version | 0.02% | 11.20 /100 | 99.94% | 99.97% |
| 90108 | wall-tile | surface | Proposed update · not adopted | 0.02% | 11.20 /100 | 99.94% | 99.97% |
| 458410 | wall-wood | surface | Original selection | 0.36% | 2.42 /100 | 86.42% | 99.64% |
| 458410 | wall-wood | surface | Earlier cleanup | 0.01% | 1.26 /100 | 99.42% | 99.99% |
| 458410 | wall-wood | surface | Cleanup with shape checks | 0.01% | 1.26 /100 | 99.42% | 99.99% |
| 458410 | wall-wood | surface | September test version | 0.01% | 1.26 /100 | 99.42% | 99.99% |
| 458410 | wall-wood | surface | Proposed update · not adopted | 0.01% | 1.26 /100 | 99.42% | 99.99% |
| 402720 | wall-wood | surface | Original selection | 95.32% | 69.00 /100 | 0.82% | 3.92% |
| 402720 | wall-wood | surface | Earlier cleanup | 95.58% | 72.96 /100 | 0.84% | 3.64% |
| 402720 | wall-wood | surface | Cleanup with shape checks | 95.58% | 72.96 /100 | 0.84% | 3.64% |
| 402720 | wall-wood | surface | September test version | 95.58% | 72.96 /100 | 0.84% | 3.64% |
| 402720 | wall-wood | surface | Proposed update · not adopted | 95.58% | 72.96 /100 | 0.84% | 3.64% |
Download the evidence
Original datasets: COCO and COCO-Stuff · COCO-Stuff label definitions · SAM 2 upstream project.
The original eight-file download contains scores, public image identifiers and source hashes. Three separately credited test photos and selection overlays can be inspected above. The Rendero showcase images are separate and have no benchmark scores. No private account data are included.
Explore AI rendering workflows · About Rendero · More research and guides