Trained on an original dataset of 1.1 million street-level urban and rural geo-tagged images, it achieves state-of-the-art performance on multiple open-domain image geolocalization benchmarks in zero-shot, outperforming supervised models trained on millions of images.

Model Description

StreetCLIP is a model pretrained by deriving image captions synthetically from image class labels using a domain-specific caption template. This allows StreetCLIP to transfer its generalized zero-shot learning capabilities to a specific domain (i.e. the domain of image geolocalization). StreetCLIP builds on the OpenAI's pretrained large version of CLIP ViT, using 14x14 pixel patches and images with a 336 pixel side length.

Model Details

Model type: CLIP
Language: English
License: Create Commons Attribution Non Commercial 4.0
Trained from model: openai/clip-vit-large-patch14-336

Model Sources

Paper: Preprint
Cite preprint as:

bibtex

  @misc{haas2023learning,
      title={Learning Generalized Zero-Shot Learners for Open-Domain Image Geolocalization}, 
      author={Lukas Haas and Silas Alberti and Michal Skreta},
      year={2023},
      eprint={2302.00275},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
  }

Uses

StreetCLIP has a deep understanding of the visual features found in street-level urban and rural scenes and knows how to relate these concepts to specific countries, regions, and cities. Given its training setup, the following use cases are recommended for StreetCLIP.

Direct Use

StreetCLIP can be used out-of-the box using zero-shot learning to infer the geolocation of images on a country, region, or city level. Given that StreetCLIP was pretrained on a dataset of street-level urban and rural images, the best performance can be expected on images from a similar distribution.

Broader direct use cases are any zero-shot image classification tasks that rely on urban and rural street-level understanding or geographical information relating visual clues to their region of origin.

Downstream Use

StreetCLIP can be finetuned for any downstream applications that require geographic or street-level urban or rural scene understanding. Examples of use cases are the following:

Understanding the Built Environment

Analyzing building quality
Building type classifcation
Building energy efficiency Classification

Analyzing Infrastructure

Analyzing road quality
Utility pole maintenance
Identifying damage from natural disasters or armed conflicts

Understanding the Natural Environment

Mapping vegetation
Vegetation classification
Soil type classifcation
Tracking deforestation

General Use Cases

Street-level image segmentation
Urban and rural scene classification
Object detection in urban or rural environments
Improving navigation and self-driving car technology

Out-of-Scope Use

Any use cases attempting to geolocate users' private images are out-of-scope and discouraged.

Bias, Risks, and Limitations

StreetCLIP was not trained on social media images or images of identifable people for a reason. As such, any use case attempting to geolocalize users' private images

Recommendations

We encourage the community to apply StreetCLIP to applications with significant social impact of which there are many. The first three categories of potential use cases under Downstream Use list potential use cases with social impact to explore.

How to Get Started with the Model

Use the code below to get started with the model.

python

from PIL import Image
import requests

from transformers import CLIPProcessor, CLIPModel

model = CLIPModel.from_pretrained("geolocal/StreetCLIP")
processor = CLIPProcessor.from_pretrained("geolocal/StreetCLIP")

url = "https://huggingface.co/geolocal/StreetCLIP/resolve/main/sanfrancisco.jpeg"
image = Image.open(requests.get(url, stream=True).raw)

choices = ["San Jose", "San Diego", "Los Angeles", "Las Vegas", "San Francisco"]
inputs = processor(text=choices, images=image, return_tensors="pt", padding=True)

outputs = model(**inputs)
logits_per_image = outputs.logits_per_image # this is the image-text similarity score
probs = logits_per_image.softmax(dim=1) # we can take the softmax to get the label probabilities

Training Details

Training Data

StreetCLIP was trained on an original, unreleased street-level dataset of 1.1 million real-world, urban and rural images. The data used to train the model comes from 101 countries, biased towards western countries and not including India and China.

Preprocessing

Same preprocessing as openai/clip-vit-large-patch14-336.

Training Procedure

StreetCLIP is initialized with OpenAI's pretrained large version of CLIP ViT and then pretrained using the synthetic caption domain-specific pretraining method described in the paper corresponding to this work. StreetCLIP was trained for 3 epochs using an AdamW optimizer with a learning rate of 1e-6 on 3 NVIDIA A100 80GB GPUs, a batch size of 32, and gradient accumulation of 12 steps.

StreetCLIP was trained with the goal of matching images in the batch with the caption correponding to the correct city, region, and country of the images' origins.

Evaluation

StreetCLIP was evaluated in zero-shot on two open-domain image geolocalization benchmarks using a technique called hierarchical linear probing. Hierarchical linear probing sequentially attempts to identify the correct country and then city of geographical image origin.

Testing Data and Metrics

Testing Data

StreetCLIP was evaluated on the following two open-domain image geolocalization benchmarks.

IM2GPS.
IM2GPS3K

Metrics

The objective of the listed benchmark datasets is to predict the images' coordinates of origin with as little deviation as possible. A common metric set forth in prior literature is called Percentage at Kilometer (% @ KM). The Percentage at Kilometer metric first calculates the distance in kilometers between the predicted coordinates to the ground truth coordinates and then looks at what percentage of error distances are below a certain kilometer threshold.

Results

IM2GPS

Model	25km	200km	750km	2,500km
PlaNet (2016)	24.5	37.6	53.6	71.3
ISNs (2018)	43.0	51.9	66.7	80.2
TransLocator (2022)	48.1	64.6	75.6	86.7
Zero-Shot CLIP (ours)	27.0	42.2	71.7	86.9
Zero-Shot StreetCLIP (ours)	28.3	45.1	74.7	88.2
Metric: Percentage at Kilometer (% @ KM)

IM2GPS3K

Model	25km	200km	750km	2,500km
PlaNet (2016)	24.8	34.3	48.4	64.6
ISNs (2018)	28.0	36.6	49.7	66.0
TransLocator (2022)	31.1	46.7	58.9	80.1
Zero-Shot CLIP (ours)	19.5	34.0	60.0	78.1
Zero-Shot StreetCLIP (ours)	22.4	37.4	61.3	80.4
Metric: Percentage at Kilometer (% @ KM)

Summary

Our experiments demonstrate that our synthetic caption pretraining method is capable of significantly improving CLIP's generalized zero-shot capabilities applied to open-domain image geolocalization while achieving state-of-the-art performance on a selection of benchmark metrics.

Environmental Impact

Hardware Type: 4 NVIDIA A100 GPUs
Hours used: 12

Citation

Cite preprint as:

bibtex

  @misc{haas2023learning,
      title={Learning Generalized Zero-Shot Learners for Open-Domain Image Geolocalization}, 
      author={Lukas Haas and Silas Alberti and Michal Skreta},
      year={2023},
      eprint={2302.00275},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
  }