EviLens

Images for EviLens, a human-verified benchmark of 688 instances for evaluating perception under insufficient evidence — cases where a single glance at the image is not enough to produce the required perceptual output.

Conventional perception benchmarks assume the image plus the model's parametric knowledge suffice to resolve the query. EviLens deliberately breaks that assumption in three ways: the target may be too small to resolve without inspecting the image at a finer scale (in many cases well under 0.1% of the image area), it may only be findable by comparing or scanning multiple regions, or identifying it may require knowledge from outside the image. The model must still commit to a box, a mask, or a count.

Composition

Category	n	Missing evidence	Metrics
Localization	140	Target is explicitly named but tiny or hidden among clutter (high-resolution scenes, I-spy puzzles)	mean box IoU, R@0.5
Recognition	182	Target is visually salient but referred to indirectly; needs external information to identify	mean box IoU, R@0.5
Spot-the-difference	15	Target is defined relative to a second panel; needs cross-panel comparison	micro-F1, macro-F1
Segmentation	195	Recognition setting, but the output is a mask	gIoU, cIoU
Counting	156	Recognition setting, but the output is an integer	exact-match accuracy

Localization, recognition, and spot-the-difference are the three grounding categories. The 15 spot-the-difference instances contain 79 annotated differences in total. Every instance is manually verified.

Metric definitions

• IoU / R@0.5 — mean box IoU, and the fraction of instances with IoU ≥ 0.5.
• gIoU / cIoU — mean per-instance mask IoU, and cumulative intersection over cumulative union. Predicted boxes are converted to masks with SAM3.
• micro-F1 / macro-F1 (spot-the-difference) — greedy one-to-one matching at IoU 0.5 between predicted boxes and annotated differences. Micro-F1 pools counts across images; macro-F1 averages per-image F1. Micro-F1 is the primary metric: with 15 images and 79 differences, it is the more stable of the two.
• Coordinates are canonically xyxy, normalized to [0, 1000].

Getting started

The evaluation harness fetches these images for you:

git clone <code repository>
cd EviLens
pip install -r requirements.txt

# downloads this dataset into benchmark/images/ and verifies every file
# against scripts/images_manifest.json (size + sha256)
python3 scripts/download_images.py

# evaluate any OpenAI-compatible endpoint
MODEL_NAME=my-model BASE_URL=https://api.example.com/v1 API_KEY=... \
  ./run_eval.sh my_run

To fetch the images on their own:

from huggingface_hub import snapshot_download
snapshot_download("bunny127/EviLens", repo_type="dataset",
                  local_dir="benchmark", allow_patterns=["images/**"])

Layout

images/
├── counting_rl/images/...                 156 files   240 MB
├── detailed_rl/{images,images_hard,...}   151 files   359 MB
└── searched/images/...                    367 files   280 MB

Each benchmark record carries image_tag and image_path; the image resolves to images/<image_tag>/<image_path>. 674 unique files serve the 688 instances — spot-the-difference pairs and a few cross-category instances share an image.

License

Data, including these images, is released under CC BY-NC 4.0: share and adapt for non-commercial purposes, with attribution.

A substantial part of this image set was collected from public web sources, including web-search results. Copyright in those images remains with their respective owners, and they are included here for non-commercial research only. If you hold rights to an image here and want it removed, open a discussion and it will be taken out of the distribution.

Citation

EviLens is introduced in EviRover: Reinforcing Agentic Perception Beyond a Glance.