# One image can become two COCO image IDs

18 November 2026. In a controlled Label Studio SDK 2.1.2 conversion, two valid box judgments for one task and one local image generated two COCO image IDs. Both image rows carried the exact same file path. Each box referenced a different generated image ID. The pinned COCO Python index then listed both IDs separately. With `LABEL_STUDIO_FORCE_ANNOTATOR_EXPORT` unset in this fixture, the box rows carried generated annotation IDs and omitted the original judgment and actor IDs. [Converter image rows](https://github.com/HumanSignal/label-studio-sdk/blob/c20c899aaaa302740dacd46a47e5774bb677561f/src/label_studio_sdk/converter/converter.py#L767-L830) · [Box rows](https://github.com/HumanSignal/label-studio-sdk/blob/c20c899aaaa302740dacd46a47e5774bb677561f/src/label_studio_sdk/converter/converter.py#L885-L936) · [COCO index](https://github.com/ppwwyyxx/cocoapi/blob/2971b0430e0990f119c60c8708ca66d4a5fed5d0/PythonAPI/pycocotools/coco.py#L89-L117) · [Unfiltered image IDs](https://github.com/ppwwyyxx/cocoapi/blob/2971b0430e0990f119c60c8708ca66d4a5fed5d0/PythonAPI/pycocotools/coco.py#L177-L196)

| Source for one image | COCO result | Unfiltered COCO index |
| --- | --- | --- |
| Two disagreeing judgments | Two rows for the same path, with result B on ID 0 and A on ID 1 | IDs 0 and 1 |
| Two agreeing judgments | Two rows for the same path, with result A on both IDs | IDs 0 and 1 |
| One selected judgment | One row for the path, with result A on ID 0 | ID 0 |

The converter emits one item per eligible judgment. It allocates an image ID and appends a new image entry for each item in the tested branch. Both agreement and disagreement yielded two IDs. The [paired YOLO fixture](README.md) sends two disagreeing judgments for one image to a single output path, where the successful older write replaces the newer one. Count distinct delivered image paths and check the source judgment-to-image mapping when a dataset is supposed to contain one row per underlying image. Unique numeric IDs alone leave that property unverified. An evaluation's score depends on how its caller uses the IDs, and no score was measured here. [Shared annotation projection](https://github.com/HumanSignal/label-studio-sdk/blob/c20c899aaaa302740dacd46a47e5774bb677561f/src/label_studio_sdk/converter/converter.py#L593-L653)

Run `python3 -B check-coco-identity.py | cmp - coco-identity-observations.json` with pipeline failure checking. The [checker](check-coco-identity.py), SHA-256 `fb6f5a0eb3b8f269576353203f40e81b46515e55182f9b377b45ba34e4925d82`, uses the exact annotation projection, rectangle geometry and COCO writer from the pinned SDK and the exact index and ID query from the pinned COCO Python source. A local image-size stand-in supplies 100 by 100 for a fixed fictional image. The receipt compares the full source image path before normalizing it to a stable basename. Two local runs and an independent replay matched all three cases byte for byte. [Saved observations](coco-identity-observations.json), SHA-256 `bb4b13aae5a7c471507f1ebcdb9cac2dcd7c2fe62f232489b0a00078ec12e222`.

An earlier report documents the older-judgment effect in YOLO. The COCO search found adjacent reports on ambiguous empty results and extra annotations from detached brush masks. This pilot checks image IDs for multiple judgments of one image. It uses public pinned Apache-2.0 SDK and BSD-licensed COCO code fetched at runtime and ships its own code and invented results. The installed app, media decoding, model training and evaluation remain outside the fixture. [Earlier YOLO report](https://github.com/HumanSignal/label-studio/issues/5421) · [Empty COCO issue](https://github.com/HumanSignal/label-studio/issues/9774) · [Brush mask issue](https://github.com/HumanSignal/label-studio/issues/7737) · [SDK license](https://github.com/HumanSignal/label-studio-sdk/blob/c20c899aaaa302740dacd46a47e5774bb677561f/LICENSE) · [COCO license](https://github.com/ppwwyyxx/cocoapi/blob/2971b0430e0990f119c60c8708ca66d4a5fed5d0/license.txt)
