Finding any Waldo with Zero-shot Invariant and Efficient Visual Search

Authors: Mengmi Zhang, Jiashi Feng, Keng Teck Ma, Joo Hwee Lim, Qi Zhao, and Gabriel Kreiman

This repository contains an implementation of a zero-shot deep learning model for invariant and efficient visual search. Our paper is published in Nature Communication.

Free access to our manuscript HERE.

Project Description

Searching for a target object in a cluttered scene constitutes a fundamental challenge in daily vision. Visual search must be selective enough to discriminate the target from distractors, invariant to changes in the appearance of the target, efficient to avoid exhaustive exploration of the image, and must generalize to locate novel target objects with zero-shot training. Previous work has focused on searching for perfect matches of a target after extensive category-specific training. Here we show for the first time that humans can efficiently and invariantly search for natural objects in complex scenes. To gain insight into the mechanisms that guide visual search, we propose a biologically inspired computational model that can selectively, invariantly and efficiently locate targets, generalizing to novel objects. The model provides an approximation to the mechanisms integrating bottom-up and top-down signals during search in natural scenes.

Stimuli	Target	Attention Map predicted by our model	Fixated Place

Pre-requisite

The code has been successfully tested in MAC OSX and Ubuntu 14.04. Only CPU is required. To speed up computation, GPU is highly recommended (3GB GPU memory at least).

It requires the deep learning platform Torch7. Refer to link for installation.

Matio package is required (save and load matlab arrays from Torch7). Refer to link for installation.

Loadcaffe package is required (load pre-trained caffe model to Torch7). Refer to link for installation.

Run the commands:

luarocks install image
luarocks install tds

Download our repository:

git clone https://github.com/kreimanlab/VisualSearchZeroShot.git

Download the caffe VGG16 model from HERE and place it in folder /Models/caffevgg16/

Usage

Visual search in object arrays

Navigate to the repository folder. To run our search model, copy the following command in the command window:

th IVSNtopdown_30_31_array.lua

Visualize the generated attention map in MATLAB: visAttentionMap_array.m

Visual search in natural images

Navigate to the repository folder and run PreprocessNaturalDesign.m in MATLAB

It preprocesses the search images by cropping it into uniform pieces and saving the paths of the cropped images in croppednaturaldesign_img.txt.

To run our search model, copy the following command in the command window:

th IVSNtopdown_30_31_naturaldesign.lua

Visualize the generated attention map in MATLAB: visAttentionMap_naturaldesign.m

Visual search in Waldo images

Navigate to the repository folder and run PreprocessWaldoImage.m in MATLAB

It preprocesses the search images by cropping it into uniform pieces and saving the paths of the cropped images in croppednaturaldesign_img.txt.

To run our search model, copy the following command in the command window:

th IVSNtopdown_30_31_waldo.lua

Visualize the generated attention map in MATLAB: visAttentionMap_waldo.m

Datasets

We have collected human eyetracking data in three increasingly complex visual search tasks: object arrays, natural images and Waldo images. Below are the preview of example trial in each task. Yellow circles denote the eye movements. Red bounding box denotes the ground truth. Correspondingly, we show the eye movement predicted by our computational model.

Object array (Human)	Natural image (Human)	Waldo image (Human)

Object array (Model)	Natural image (Model)	Waldo image (Model)

Note that we also include one variation of the search task in object arrays. Instead of familiar objects, we collect novel objects from [1] [2] [3] and create a dataset of novel objects on arrays. See below for novel object examples:

You can download the complete dataset including the novel object dataset from Part1, Part2 and Part3 (total size: ~9GB)

It contains the following:

• Part 1: Datasets folder: contain search images, targets, ground truth, psychophysics (human eyetracking data, MATLAB function to process and extract fixations)
• Part 2: Eval folder: contain MATLAB files to evaluate cummulative search performance as a function of number of fixations for our computational model across three search tasks
• Part 2: Plot folder: contain MATLAB files to reproduce figures in our paper
• Part 2: supportingFunc folder: add this directory in your MATLAB search path
• Part 3: DataForPlot Folder: pre-processed results saved in .mat. Place this folder in Plot folder (Part 2) before using plotting functions

Notes

The source code is for illustration purpose only. Path reconfigurations may be needed to run some MATLAB scripts. We do not provide techinical supports but we would be happy to discuss about SCIENCE!

License

See Kreiman lab for license agreements before downloading and using our source codes and datasets.