IC-933DNN object detectors do not exhibit the same sensitivity to image clutter as humans in peripheral object detection

Anne Harrington, Vasha DuTell, Mark Hamilton, Ayush Tewari, Simon Stent, William T. Freeman, Ruth Rosenholtz

SourceCOCO-Periph: Bridging the Gap Between Human and Machine Perception in the Periphery

Human critical eccentricity for peripheral object detection decreases as the number of nearby objects (clutter) increases, consistent with the visual crowding literature. In contrast, the tested DNN object detectors show no strong relationship between clutter and critical eccentricity; their critical eccentricities remain low regardless of the number of nearby objects. This holds even for models trained on COCO-Periph, which should in principle reflect the degrading effect of clutter on the peripheral representation.

Evidence
correlational
Caveat
Clutter is proxied by the number of ground-truth COCO annotations in the image, which does not label all objects in many scenes; the relationship is assessed qualitatively from Figure 6b without a reported correlation coefficient.
Model
DINO-FocalNet-Large, Swin Transformer DINO-Swin-Tiny, DETR-R50, RetinaNet-R50, FoveaBox, Faster R-CNN / Faster R-CNN R50 / Faster R-CNN X101
Datasets
MS COCO / COCO / COCO 2014 / COCO 2017 / COCO 20k / COCO-it / COCO-wl [source]
Methods
Uniform TTM [primary]
Related findings
IC-932, IC-934
Extraction
automatic-extraction