Multimodal Large Language Models as Image Classifiers
Paper • 2603.06578 • Published
Datasets:
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is released for research purposes only. By requesting access, you agree to: (1) cite the associated paper in any publication or project that uses this dataset; (2) not redistribute the dataset or its annotations without permission from the authors.
Log in or Sign Up to review the conditions and access this dataset content.
Note for reviewers: This dataset uses Hugging Face gated access with automatic approval to comply with NeurIPS submission requirements. Unfortunately, Hugging Face automatically shares your username and email with the dataset authors upon access request — we cannot disable this. To preserve anonymity during review, please either create a new anonymous Hugging Face account to request access, or use the copy of the dataset included in the supplementary material of our submission, which requires no authentication.
ReImageNet
ReImageNet is a complete multilabel reannotation with localization of the ImageNet-1K validation set (ILSVRC2012). A team of 7 trained in-house annotators reviewed all 50,000 validation images through an iterative annotation process, producing per-image bounding boxes with class labels and annotation attributes, correcting and extending the original single-label ground truth. Class names and definitions were revised where the original WordNet-based names no longer matched the actual image content.
Note on images: This repository contains annotations only. The images are part of the ImageNet dataset and must be obtained separately from image-net.org. Match images to annotations using the
file_pathfield (e.g.n01440764/ILSVRC2012_val_00000293.JPEG).
Dataset Summary
| Property | Value |
|---|---|
| Base dataset | ImageNet-1K validation set (ILSVRC2012) |
| Images | 50,000 |
| Bounding boxes | 99,534 |
| ImageNet classes | 1,000 |
| Labels per image | mean 1.63, median 1 |
| Bounding boxes per image | mean 1.99, median 1 |
| Single-label images (S) | 62.6% |
| Multi-label images (M) | 32.7% |
| No valid label images (N) | 4.7% |
| Annotators | 7 trained non-domain-experts |
| License (annotations) | CC BY 4.0 |
Dataset Structure
Files
| File | Description |
|---|---|
reannotation.jsonl |
Main annotation file — one JSON record per line |
label_names.json |
List of 1,000 ImageNet synset IDs indexed by class integer (0–999) |
class_update_config.json |
Configuration file containing equivalent classes (visually |
| indistinguishable ImageNet class pairs treated as interchangeable during | |
| evaluation). |
Data Fields
Each line in reannotation.jsonl is a JSON object with the following fields:
| Field | Type | Description |
|---|---|---|
image_name |
string |
Filename (e.g. ILSVRC2012_val_00004410.JPEG) |
original_class |
int[] |
Original ImageNet label(s). Usually a single integer; some images have two labels due to ImageNet equivalent classes |
reannotated_labels |
int[] |
All class labels visible in the image, as determined by annotators |
file_path |
string |
Relative path within the ImageNet val directory ({synset}/{image_name}) |
bboxes |
object[] |
List of bounding box annotations (see below) |
All class integers are indices into label_names.json (0-indexed).
Bounding Box Fields
Each element of bboxes is:
| Field | Type | Description |
|---|---|---|
coordinates |
int[4] |
Bounding box in pixel space: [x1, y1, x2, y2] (top-left, bottom-right). Boxes enclose the object with a moderate margin. |
labels |
int[] |
Class label(s) for this box (index into label_names.json) |
group |
int | null |
Group ID. When multiple bbox entries share the same non-null group value, they represent the same physical object with multiple labels. Coordinates are identical across grouped entries. |
crowd_flag |
bool |
True if the bbox covers five or more instances of the same class collapsed into a single box. Exact instance count is not recorded. |
reflected_flag |
bool |
True if the object is a reflection in a mirror or water surface, annotated independently of whether the reflected object itself is also visible. |
rendition_flag |
bool |
True if the object is an artificial or stylized representation (toy, drawing, sculpture, logo, etc.) rather than a real instance. |
ocr_needed_flag |
bool |
True if correct classification requires reading text visible in the image. |
dominant_object |
bool |
True if the object is one a person would notice immediately upon viewing the image. Based on annotator judgment rather than size alone. |
Label interpretation note: An empty
reannotated_labelslist ([]) indicates the image contains no valid ImageNet class — corresponding to theNcategory in the paper, where annotators were confident no ImageNet object is present. A label with value-1indicates the annotator was uncertain about the correct label (marked as Not Sure during annotation); these images are flagged for verification in the ongoing verififcation phase.
label_names.json
A JSON array of 1,000 WordNet synset IDs ordered by class index:
["n01440764", "n01443537", ..., "n15075141"]
label_names[i] is the synset for class integer i.
class_update_config.json
Contains two entries. eq_classes lists pairs of ImageNet classes that
are visually indistinguishable or semantically equivalent (e.g.
laptop/notebook computer, bathtub/tub), and are therefore treated
as interchangeable during evaluation — a prediction of either class is
counted as correct. metadata records the creation date and the date of
the last update of this list.
Evaluation code
Evaluation code is available at github.com/klarajanouskova/ImageNet
Annotation Attributes
Each bounding box can carry one or more attributes that capture visually distinct properties of the depicted object:
- Rendition — the object is a toy, drawing, sculpture, or other artificial representation. Tests model robustness to non-real-world depictions.
- Crowd — five or more instances of the same class, collapsed into a single bounding box. The exact count is not recorded.
- Text-recognition — correct classification requires reading visible text. Tests whether models can leverage text as a recognition cue.
- Reflection — the object appears as a reflection in a mirror or water surface, annotated regardless of whether the original object is also visible.
- Dominant — the object would be immediately noticed upon viewing the image. An image may have any number of dominant objects, or none at all.
Annotation Process
Annotations were produced by a team of 7 non-domain-expert annotators (aged 16–50, spanning diverse backgrounds including geology specialists, canine enthusiasts, and automotive buffs) recruited and trained in-house.
The annotation process was iterative:
- Training: Annotators studied known ImageNet issues, worked through attribute examples, and practised on intentionally challenging classes. Only annotators passing a quality threshold proceeded to real tasks.
- Per-class preparation: Before labelling any image, annotators examined the actual image content of each class, consulted external references (Wikipedia, iNaturalist), and recorded a working definition in a shared table.
- Image labelling: Annotators drew bounding boxes around all objects a person would notice at first glance, assigned class labels and attributes, assisted by OWLv2 bounding box proposals and OpenCLIP top-20 class predictions.
- Verification round: Annotators revisited already-annotated images using finalised guidelines, with OWLv2/OpenCLIP replaced by MLLM predictions with SAM3-generated localisations as a stronger reference. This process is ongoing.
Quality was continuously monitored via control sets and a supervised communication channel.
Considerations for Using the Data
Limitations
- All annotators share a European background (Czechia, Ukraine, Greece, Latvia), which may affect interpretation of culturally specific classes.
- Fine-grained wildlife classes were annotated by non-experts using online references; species-level distinctions may be imprecise.
- The shared class definition table enforces consistency but may propagate errors if a class is defined incorrectly.
- Model predictions (anonymised, optional) may have nudged annotators toward certain labels.
- This reannotation covers only the validation set. Due to the distribution shift between training and validation sets, revised class names and definitions may not accurately reflect the training set.
- The verification round is ongoing; some annotations may still be updated.
Social Impact
This dataset extends ImageNet-1k validation labels to support multilabel evaluation and spatial grounding, enabling more accurate measurement of model performance. The attribute annotations allow fine-grained analysis of model capabilities across distinct recognition regimes (text-based, rendition-based, etc.).
Related Work
- Flaws of ImageNet — our prior analysis of ImageNet issues, arXiv, ICLR blogpost
- Multimodal Large Language Models as Image Classifiers — partial reannotation and MLLM evaluation study
- Aiming for Perfect ImageNet-1K - project page
Dataset Card Contact
- Downloads last month
- 24