ODOR Dataset Benchmark

• ODOR dataset is a benchmark for detecting olfactory references in artworks, providing 4,712 images with 38,116 fine-grained object annotations across 139 categories.
• It poses unique challenges such as artistic abstraction, dense overlap, peripheral object placement, and a long-tailed category distribution.
• Baseline evaluations reveal modest AP scores, highlighting the difficulties in small object detection and fine-grained subclass discrimination.

Searching arXiv for the ODOR dataset and related papers to ground the article in the latest literature. The ODOR dataset—short for Object Detection for Olfactory References—is a computer-vision benchmark for fine-grained, instance-level detection of objects in artworks that carry olfactory meaning, including flowers, fruits, vessels, smoke, and related motifs. In its broader release, it contains 4,712 artwork images, 38,116 object-level bounding-box annotations, and 139 fine-grained categories, and is explicitly designed to expose challenges that are atypical in photographic detection benchmarks: artistic abstraction, dense and overlapping instances, peripheral placement, small objects, and a long-tailed category distribution (Zinnen et al., 11 Jul 2025).

1. Historical emergence and benchmark lineage

ODOR emerged within the Odeuropa research line, which frames olfactory references in European visual heritage as a detection problem. An earlier benchmark was released for the ICPR2022 ODeuropa Challenge on Olfactory Object Recognition, where the public portion comprised 2,647 paintings, prints, and drawings with 20,120 bounding boxes, and a private test set contained 1,140 artworks with 15,480 annotations (Zinnen et al., 2023). That challenge version used 87 leaf-node classes under 13 supercategories and covered historical material dating roughly from c. 1500–1900.

The later ODOR release broadens the scope to 139 fine-grained categories organized into eight supercategories and emphasizes statistical analysis of the dataset as a detection benchmark in its own right. The motivation is not only olfactory heritage, but also the need for models that remain robust under artistic style variation, subtle fine-grained distinctions, and off-center object placement in artworks (Zinnen et al., 11 Jul 2025).

This continuity is methodologically important. The challenge release established the task setting, COCO-style evaluation, and long-tail difficulty; the later release systematized these issues through a larger annotation set and more extensive baseline and secondary studies. A plausible implication is that ODOR should be read less as a single static corpus than as a benchmark family centered on olfactory object recognition in historical art.

2. Data model, taxonomy, and annotation procedure

At the dataset level, ODOR provides high-resolution artwork images together with object-level bounding boxes in standard COCO JSON. Each annotation includes at least bbox, category_id, and supercategory. The final taxonomy was obtained from an original candidate pool of 227 classes, which was pruned to 139 by requiring at least 16 instances in at least 3 images for each retained category (Zinnen et al., 11 Jul 2025).

The category system is explicitly fine-grained. Reported supercategory examples include:

Supercategory example	Subclasses	Instances
flower	30	10,866
fruit	17	7,818
mammal	14	4,972
smoking equipment	8	2,595
other	36 rare subclasses	6,356

The remaining supercategories named in the release are vegetable, drinking vessel, and smoke-related. The annotation target is not generic object presence, but olfactory relevance in artworks, which explains the prominence of florals, fruits, smoking implements, and vessels.

Annotation quality control combines crowdsourcing and expert review. Annotations were collected via Amazon Mechanical Turk and then verified and corrected in multiple rounds by art-history experts to ensure both species-level and historically accurate labeling (Zinnen et al., 11 Jul 2025). In the earlier challenge release, annotators used CVAT, with Amazon Mechanical Turk specifically mentioned for flower subtypes, and were instructed to draw “tightly fitted” boxes around every visible instance, including very small and partially occluded objects (Zinnen et al., 2023).

3. Statistical structure and sources of difficulty

ODOR is statistically challenging in several distinct senses. The mean number of categories per image is reported as $\mu_{(\mathrm{cls/img})}=3.7$, and the mean number of instances per image as $\mu_{(\mathrm{b/img})}=8.1$. The class distribution is long-tailed, and some images contain more than 15 instances or more than 8 classes (Zinnen et al., 11 Jul 2025).

A central difficulty is overlap and occlusion. Over 50% of instances overlap with at least one other instance, and many share 10–40% of their area with neighboring objects. This is paired with a marked small-object regime: a large fraction of objects occupy less than 4% of the image area, whereas only a few exceed 25% (Zinnen et al., 11 Jul 2025). Such geometry stresses localization, NMS behavior, and fine-grained classification under limited pixel support.

Another distinctive property is spatial distribution. Unlike COCO or PASCAL, which are described as having a strong center bias, ODOR objects are reported as almost uniformly distributed across the canvas. The paper attributes this to the inclusion of peripheral or background objects such as small flowers, fruits, and smoke (Zinnen et al., 11 Jul 2025). This suggests that architectural priors tuned to center-biased natural-image corpora may transfer imperfectly.

The class distribution’s complexity can, as noted in the paper, be expressed via Shannon entropy,

$H=-\sum_k p_k \log p_k,$

where $p_k$ is the fraction of instances in category $k$; however, the paper does not report an explicit value for $H$ (Zinnen et al., 11 Jul 2025). The absence of a reported entropy does not diminish the practical observation that ODOR combines long-tail imbalance with dense clutter and peripheral placement, which is an unusual conjunction among detection benchmarks.

4. Evaluation protocol and baseline detection performance

ODOR adopts standard COCO-style detection metrics. The reported evaluation uses mean Average Precision over IoU thresholds 0.50:0.05:0.95, together with AP$_{50}$, AP$_{75}$, and size-stratified AP$_s$, AP$_m$, and AP$\mu_{(\mathrm{b/img})}=8.1$0. The paper defines

$\mu_{(\mathrm{b/img})}=8.1$1

where $\mu_{(\mathrm{b/img})}=8.1$2 is precision as a function of recall for class $\mu_{(\mathrm{b/img})}=8.1$3 (Zinnen et al., 11 Jul 2025).

The benchmark spans two-stage, transformer-based, one-stage anchor-free, and hybrid detectors. Reported test results include the following:

Model	AP	AP$\mu_{(\mathrm{b/img})}=8.1$4
F-RCNN (ResNet-50)	10.1	19.3
DINO (ResNet-50)	13.3	22.2
MADet (ResNet-50)	10.2	18.7
FCOS (ResNet-50)	8.0	14.3
YOLO-v8L	18.5	29.0
DINO + FocalNet-L	22.6	36.5

Within Faster R-CNN, stronger backbones improve performance up to AP = 19.0 with SWIN-L. The best reported system overall is DINO + FocalNet-L, with AP = 22.6, AP$\mu_{(\mathrm{b/img})}=8.1$5 = 36.5, and AP$\mu_{(\mathrm{b/img})}=8.1$6 = 23.5 (Zinnen et al., 11 Jul 2025).

The secondary analyses clarify why these numbers remain modest. When predictions are evaluated only at the supercategory level, AP increases by 35–60% across methods; for DINO + FocalNet-L, subclass AP = 22.6% rises to supercategory AP = 32.5%. This isolates fine-grained subclass discrimination—not mere localization—as a major bottleneck. Size-based evaluation reaches approximately AP$\mu_{(\mathrm{b/img})}=8.1$7 = 12.2%, AP$\mu_{(\mathrm{b/img})}=8.1$8 = 24.3%, and AP$\mu_{(\mathrm{b/img})}=8.1$9 = 40.3%, confirming that small-object detection is the principal failure mode (Zinnen et al., 11 Jul 2025).

Class-wise results further show that frequency alone does not determine difficulty. Lizard with 31 training instances and average size 22,355 px$H=-\sum_k p_k \log p_k,$0 attains AP = 67.4%, whereas grapes, despite 1,576 training instances and average size 7,086 px$H=-\sum_k p_k \log p_k,$1, achieve only AP = 13.8% because of occlusion and background blending. “Jan Steen Jug” with 37 instances and average size 40,234 px$H=-\sum_k p_k \log p_k,$2 reaches AP = 63.0% (Zinnen et al., 11 Jul 2025). The implied lesson is that object scale, visual separability, and contextual crowding can dominate raw sample count.

5. Research uses, methodological lessons, and recommended practices

ODOR is positioned at the intersection of digital humanities, cultural heritage computing, and computer vision. Concrete use cases listed for the dataset include quantitative art history, such as tracing the appearance of smoking vessels or roses over time; iconographic analysis of historically significant objects like tobacco pipes or incense burners; and “material turn” studies concerned with the social life of things, sensory history, and microhistory (Zinnen et al., 11 Jul 2025).

For machine learning, ODOR is described as a mid-sized, richly annotated resource that can support multi-stage fine-tuning from natural-image datasets toward artistic domains. The available metadata—artist, year, Iconclass codes, and image titles/descriptions—can be exploited for multimodal or multi-task learning, including CLIP fine-tuning and image-text co-training (Zinnen et al., 11 Jul 2025). This suggests that ODOR is not merely a detector benchmark but also a substrate for representation learning on visually and historically structured cultural material.

The earlier challenge documentation identifies several methodological directions that are especially compatible with ODOR’s failure modes: domain adaptation from photo-trained backbones, hierarchical detection that can back off to supercategory labels when fine-grained confidence is low, and few-shot learning or re-balancing losses such as focal loss and class-balanced sampling for tail classes (Zinnen et al., 2023). It also reports that the dominant recipe in top-ranked challenge systems combined pre-training on ImageNet or COCO with fine-tuning at moderate learning rates, while mild geometric and color augmentations were sufficient and more exotic augmentations did not consistently help.

Future directions named for the later dataset include semi-automated annotation to enlarge the current image base, inclusion of additional textual metadata, more robust URL hosting, and future ODOR challenges that would push toward instance segmentation and even rarer classes (Zinnen et al., 11 Jul 2025).

6. Terminological ambiguity: ODOR in other olfactory machine-learning literatures

The label “ODOR dataset” is not unique to artwork object detection, and disambiguation is often necessary in arXiv-era literature. In cheminformatics, “ODOR” has also been used for an ordinal odor strength dataset containing 2,056 molecules labeled into odorless, low, medium, and high categories, assembled from Good Scents and PubChem and released for structure-based prediction of odor intensity (Fichtelmann et al., 9 Dec 2025).

A separate molecular resource associated with odor taxonomies aggregates 6,711 unique molecules annotated with 146 distinct odor descriptors, together with an Expert Taxonomy and a Data-driven Taxonomy; its files include MMD.csv, descriptors_matrix.csv, ET_hierarchy.json, and DT_partition.csv (Sajan et al., 11 Aug 2025).

ODOR usage	Domain	Scale
Object Detection for Olfactory References	artwork object detection	4,712 images; 38,116 annotations; 139 categories
Ordinal odor strength dataset	molecular odor-strength prediction	2,056 molecules; 4 ordinal bins
Odor taxonomy dataset	structure-based odor prediction	6,711 molecules; 146 descriptors

For scholarship on visual cultural heritage, ODOR ordinarily denotes the Object Detection for Olfactory References benchmark (Zinnen et al., 11 Jul 2025). In molecular olfaction, however, the same label can refer to chemically grounded datasets with entirely different units of analysis, target variables, and evaluation protocols (Fichtelmann et al., 9 Dec 2025, Sajan et al., 11 Aug 2025). This naming overlap is a practical source of citation ambiguity and should be resolved explicitly in technical writing.