Visual perspective taking in humans and machines

Drew Linsley¹, Peisen Zhou¹, Alekh Karkada¹, Akash Nagaraj¹, Gaurav Gaonkar¹, Francis E Lewis¹, Zygmunt Pizio², Thomas Serre¹

¹Carney Institute for Brain Science, Brown University, Providence, RI.

²Department of Cognitive Sciences, University of California-Irvine, Irvine, CA.

Project Page · Paper · Data

Paper Abstract

Visual perspective taking (VPT), the ability to accurately perceive and reason about the perspectives of others, is an essential feature of human intelligence. VPT is a byproduct of capabilities for analyzing 3-Dimensional (3D) scenes, which develop over the first decade of life. Deep neural networks (DNNs) may be a good candidate for modeling VPT and its computational demands in light of a growing number of reports indicating that DNNs gain the ability to analyze 3D scenes after training on large static-image datasets. Here, we investigated this possibility by developing the 3D perception challenge (3D-PC) for comparing 3D perceptual capabilities in humans and DNNs. We tested over 30 human participants and ''linearly probed'' or text-prompted over 300 DNNs on several 3D-analysis tasks posed within natural scene images: (i.) a simple test of object depth order, (ii.) a basic VPT task (VPT-basic), and (iii.) a more challenging version of VPT designed to limit the effectiveness of ''shortcut'' visual strategies. Nearly all of the DNNs approached or exceeded human accuracy in analyzing object depth order, and surprisingly, their accuracy on this task correlated with their object recognition performance. In contrast, there was an extraordinary gap between DNNs and humans on VPT-basic: humans were nearly perfect, whereas most DNNs were near chance. Fine-tuning DNNs on VPT-basic brought them close to human performance, but they, unlike humans, dropped back to chance when tested on VPT-perturbation. Our challenge demonstrates that the training routines and architectures of today's DNN are well-suited for learning basic 3D properties of scenes and objects, but not for reasoning about these properties in the way that humans rely on for their everyday lives. We release our 3D-PC datasets and code to help bridge this gap in 3D perception between humans and machines.

Data Access

Hugging Face

We release data for all three tasks: VPT-basic, VPT-strategy, and depth order on Hugging Face.

https://huggingface.co/datasets/pzhou10/3D-PC

from datasets import load_dataset
# config_name: one of ["vpt-basic", "vpt-strategy", "depth"]
dataset = load_dataset("pzhou10/3D-PC", "vpt-basic")

Download the full dataset

We release the complete 3D-PC dataset along with data splits for training and testing.

https://connectomics.clps.brown.edu/tf_records/VPT/

Dataset Content

train contains all training images organized by categories.

train
|
|_<category>
|  |_<object>
|    |_<setting>
|      |_<*.png>

The corresponding labels are train_perspective.csv and depth_perspective.csv. We also provide train_perspective_balanced.csv and depth_perspective_balanced.csv, where the numbers of positive and negative samples are equal.

perspective and depth contain all data splits for 'VPT' and 'depth order' tasks.

perspective/depth
|
|_<split>
|  |_<category> 0/1
|    |_<*.png>

TIMM Evaluation

To linear probe a timm model

python run_linear_probe.py --task <task> --data_dir <data_folder>/<task>/ --model_name <model_name>

To fine-tune a timm model

python run_finetune.py --task <task> --data_dir <data_folder>/<task>/ --model_name <model_name>

data_folder: Root directory for the dataset

task: Either perspective or depth

model_name: TIMM model name

Citation

@misc{linsley20243dpc,
      title={The 3D-PC: a benchmark for visual perspective taking in humans and machines}, 
      author={Drew Linsley and Peisen Zhou and Alekh Karkada Ashok and Akash Nagaraj and Gaurav Gaonkar and Francis E Lewis and Zygmunt Pizlo and Thomas Serre},
      year={2024},
      eprint={2406.04138},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}