Every object in a scene, hand-segmented — including surfaces like grass, walls, and sky.
Mapping fixations to objects segmented from cluttered, real-world scenes
paper pdf poster colab github elizabethhhall.com
Every object in a scene, hand-segmented — including surfaces like grass, walls, and sky.
Objects in Focus is a suite of data and tools for mapping eye fixations onto objects in cluttered, real-world scenes. Scene-viewing research has long relied on pixel-level saliency, but people look at things—so the unit of analysis should be the object. Getting there requires segmentations that are exhaustive rather than convenient: not just the handful of nameable foreground items, but every surface and background region a fixation might land on.
This project documents that segmentation process across 100 scenes, releases the resulting masks and annotations alongside depth estimates, and provides notebooks for assigning fixations to objects. The accompanying paper reports that image size and viewing perspective have a substantial effect on how fixations distribute across objects.
| images/ | The 100 source scene photographs, full resolution. |
| masks/ | Segmentation masks as NumPy arrays, one per scene. |
| annotations/ | CVAT-format XML with object polygons and labels. |
| depth/ | Monocular depth (disparity) estimates for each scene. |
| notebooks/ | Colab-ready notebooks for mapping fixations to objects, sorting by depth, generating COCO-style annotations, and running DeepGaze comparisons. |
| papers/ | The ICCV 2025 HiCV paper and conference poster. |
| figures/ | Figures from the paper. |
Start with the Objects in Focus notebook for segmentation and fixation-mapping functions.
Scenes paired with object-level maps: once fixations are assigned to objects, attention can be summarized per object rather than per pixel.
Objects per scene. Most scenes contain 15–35 segmented objects; the most cluttered contains 74.
Fixation positions grouped into spatial clusters within a scene.
Paper on the segmentation process
Objects in focus: How object spatial probability underscores eye movement patterns
Elizabeth H. Hall, Zoe Loh — ICCV 2025, Humans in Computer Vision
Papers using these segmentations
Looking for Semantic Similarity
Object Selection in Passive Viewing
Surfaces and Visual Search
All 100 scenes in the set. Click any image for the full-resolution version.
