Exploring Visual Engagement Signals for Representation Learning

Menglin Jia, Zuxuan Wu, Austin Reiter, Claire Cardie, Serge Belongie and Ser-Nam Lim
Cornell University, Facebook AI

teaser-3

teaser_small

VisE is a pretraining approach which leverages Visual Engagement clues as supervisory signals. Given the same image, visual engagement provide semantically and contextually richer information than conventional recognition and captioning tasks. VisE transfers well to subjective downstream computer vision tasks like emotion recognition or political bias classification.

? Loading pretrained models

:exclamation:NOTE: This is a torchvision-like model (all the layers before the last global average-pooling layer.). Given a batch of image tensors with size (B, 3, 224, 224), the provided models produce spatial image features of shape (B, 2048, 7, 7), where B is the batch size.

Loading models with torch.hub

Get the pretrained ResNet-50 models from VisE in one line!

VisE-250M (ResNet-50): this model is pretrained with 250 million public image posts.

import torch
model = torch.hub.load("KMnP/vise", "resnet50_250m", pretrained=True)

VisE-1.2M (ResNet-50): This model is pretrained with 1.23 million public image posts.

import torch
model = torch.hub.load("KMnP/vise", "resnet50_1m", pretrained=True)

Loading models manually

	Arch	Size	Model
VisE-250M	ResNet-50	94.3 MB	download
VisE-1.2M	ResNet-50	94.3 MB	download

If you encounter any issues with torch.hub, alternatively you can download the model checkpoints manually, and then following the script below.

import torch
import torchvision

# Create a torchvision resnet50 with randomly initialized weights.
model = torchvision.models.resnet50(pretrained=False)

# Get the model before the global aver-pooling layer.
model = torch.nn.Sequential(*list(model.children())[:-2])

# load the pretrained model from a local path: <CHECKPOINT_PATH>:
model.load_state_dict(torch.load(CHECKPOINT_PATH))

? Citing VisE

If you find VisE useful in your research, please cite the following publication.

@misc{jia2021vise,
      title={Exploring Visual Engagement Signals for Representation Learning}, 
      author={Menglin Jia and Zuxuan Wu and Austin Reiter and Claire Cardie and Serge Belongie and Ser-Nam Lim},
      year={2021},
      eprint={2104.07767},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}