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

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

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate this by (1) modifying the texture tiling model (TTM), a well tested model of peripheral vision, to be more flexibly used with DNNs, (2) generating a large dataset which we call COCO-Periph that contains images transformed to capture the information available in human peripheral vision, and (3) comparing DNNs to humans at peripheral object detection using a psychophysics experiment.
Researcher Affiliation Collaboration 1MIT CSAIL 2MIT Brain and Cognitive Sciences 3 Toyota Research Institute
Pseudocode Yes We provide pseudo-code for the machine psychophysics procedure. Algorithm 1 For each object present/absent image in the human experiment, we create 100 pairings of uniform TTM transform images (P and A). We simulate trials by looking at the box predictions (boxes) of a detection DNN for each pairing (p, a).
Open Source Code Yes We publicly release our COCO-Periph dataset, along with code for uniform TTM and the psychophysics analyses at https://github.com/Rosenholtz Lab/COCOPeriph to enable further research into human and machine perception in the periphery paving the way for DNNs to mimic and benefit from properties of human visual processing.
Open Datasets Yes We apply Uniform TTM to the COCO dataset, creating COCO-Periph which contains images transformed like peripheral vision. [...] COCO-Periph contains the entire COCO 2017 validation and test set, and over 74K, 117K, 118K, and 97K of the 118K total training images transformed to model 5 , 10 , 15 , and 20 of eccentricity, respectively.
Dataset Splits Yes COCO-Periph contains the entire COCO 2017 validation and test set
Hardware Specification No The paper mentions using 'MIT Super Cloud' and 'Lincoln Laboratory Supercomputing Center' for 'HPC resources' but does not specify exact hardware components like GPU/CPU models or specific configurations.
Software Dependencies No The paper mentions using Detectron2, the La Ma image in-painting model, Eye Link 1000, and the detrex library, but does not provide specific version numbers for these software components.
Experiment Setup Yes All models are trained for 180,000 iterations starting from the weights of a pretrained R-CNN from (Wu et al., 2019). We set the solver to step at 120,000 and 160,000. We set the base learning rate to 3 10 4. All other training parameters are the same R-CNN training parameters in (Wu et al., 2019) as the baseline model.