IC-932Pre-trained DNN object detectors show a sharp falloff in peripheral detection performance with increasing eccentricity, degrading to near-chance by 20°, while human performance degrades gradually
Anne Harrington, Vasha DuTell, Mark Hamilton, Ayush Tewari, Simon Stent, William T. Freeman, Ruth Rosenholtz
Seven pre-trained object detection models were evaluated on a two-interval forced-choice peripheral detection task using uniform TTM-transformed COCO images at 5, 10, 15, and 20 degrees eccentricity. All models show a steep drop in average precision and critical eccentricity as eccentricity increases, reaching near-chance performance by 20 degrees. In contrast, 10 human subjects show a gradual, graceful decline in detection accuracy across the same eccentricities. Humans outperform all DNNs, with critical eccentricity thresholds more than 5 degrees greater than any detection model.
Evidence
correlational
Key metric
AP at 0°/5°/10°/15°/20°: DINO-FocalNet-Large 58.4/51.6/44.4/20.2/15.0; DINO-Swin-Tiny 51.3/44.0/34.1/11.1/7.6; DETR-R50 42.0/35.2/25.1/6.9/4.5; RetinaNet-R50 38.7/31.5/22.1/6.9/5.0; FoveaBox 40.4/33.2/23.4/7.5/5.3; Faster R-CNN X101 39.6/32.6/21.8/6.6/4.7; Faster R-CNN R50 36.7/29.4/19.9/5.9/4.2. Human critical eccentricity exceeds all DNNs by more than 5°.
Caveat
TTM may under-predict human performance and has been primarily validated on greyscale images; the machine psychophysics scoring protocol involves design choices (padded boxes, summed scores, no label enforcement) that may affect absolute performance levels.