doppelgangers

Doppelgangers: Learning to Disambiguate Images of Similar Structures

Ruojin Cai¹, Joseph Tung¹, Qianqian Wang¹, Hadar Averbuch-Elor², Bharath Hariharan¹, Noah Snavely^1
1Cornell University, ²Tel Aviv University
ICCV 2023 Oral

Project Page | Paper | Arxiv | Dataset

Implementation for paper “Doppelgangers: Learning to Disambiguate Images of Similar Structures”, which proposes a learning-based approach to disambiguate distinct yet visually similar image pairs (doppelgangers) and applies it to structure-from-motion disambiguation.

Dependencies

1. Create Conda environment with Python 3.8, PyTorch 1.12.0, and CUDA 10.2 using the following commands:

conda env create -f ./environment.yml
conda activate doppelgangers

2. Download Colmap from their installation page (version 3.8+).

Pretrained Model

• Pretrained model is available in the following link. To use the pretrained models, download the checkpoint.tar.gz file, unzip and put doppelgangers_classifier_loftr.pt under ./weights/ folder.

  [Click to expand]

  cd weights/
  wget -c https://doppelgangers.cs.cornell.edu/dataset/checkpoint.tar.gz
  tar -xf checkpoint.tar.gz
  mv doppelgangers/checkpoints/doppelgangers_classifier_loftr.pt ./
  rm checkpoint.tar.gz
  rm -r doppelgangers/
  

• We use LoFTR models for feature matching. Please download the LoFTR outdoor checkpoint outdoor_ds.ckpt in the following Google Drive link, and put it under ./weights/ folder.

Dataset

This section contains download links for several helpful datasets:

• SfM Disambiguation: Download the Structure from Motion disambigaution Dataset from the SfM disambiguation with COLMAP GitHub repository. Unzip the file under folder ./data/sfm_disambiguation/.
• Pairwise Visual Disambiguation: Download the Doppelgangers Dataset by following these instructions and put the dataset under folder ./data/doppelgangers_dataset/.

SfM Disambiguation

Overview of applying Doppelgangers classifier to SfM disambiguation with COLMAP

• We begin by performing COLMAP feature extraction and matching, which generates a list of image pairs.
• Next, we run Doppelgangers classifiers on these image pairs, with LoFTR matches as input.
• We then remove image pairs from the COLMAP database if they have a predicted probability below the specified threshold. These pairs are more likely to be Doppelgangers.
• Finally, we perform COLMAP reconstruction with the pruned database.

Demo

We provide a demo on the Cup dataset to demonstrate how to use our Doppelgangers classifier in Structure from Motion disambiguation with COLMAP: ./notebook/demo_sfm_disambiguation.ipynb.

Script

We provide a script for SfM disambiguation. The COLMAP reconstruction with Doppelgangers classifier can be found at [output_path]/sparse_doppelgangers_[threshold]/.

# Usage:
# python script_sfm_disambiguation.py [path/to/example_config] --input_image_path [path/to/dataset] --output_path [path/to/output]
python script_sfm_disambiguation.py doppelgangers/configs/test_configs/sfm_disambiguation_example.yaml \
  --colmap_exe_command colmap \
  --input_image_path data/sfm_disambiguation/yan2017/cup/images \
  --output_path results/cup/ \
  --threshold 0.8 \
  --pretrained weights/doppelgangers_classifier_loftr.pt

Reconstruction Results

COLMAP reconstructions of test scenes with and without the Doppelgangers classifier, as described in the paper: reconstructions.tar.gz (3G)

These reconstructions can be imported and visualized in the COLMAP GUI.

Pairwise Visual Disambiguation

Method overview

(a) Given a pair of images, we extract keypoints and matches via feature matching methods. Note that this is a negative (doppelganger) pair picturing opposite sides of the Arc de Triomphe. The feature matches are primarily in the top part of the structure, where there are repeated elements, as opposed to the sculptures on the bottom part. (b) We create binary masks of keypoints and matches. We then align the image pair and masks with an affine transformation estimated from matches. © Our classifier takes the concatenation of the images and binary masks as input and outputs the probability that the given pair is positive.

Testing the pretrained model

# Usage:
# python test.py [path/to/config] --pretrained [path/to/checkpoint]
python test.py doppelgangers/configs/training_configs/doppelgangers_classifier_noflip.yaml \
  --pretrained weights/doppelgangers_classifier_loftr.pt

Training

# Usage:
# python train.py [path/to/config] 
# python train_multi_gpu.py [path/to/config]
python train.py doppelgangers/configs/training_configs/doppelgangers_classifier_noflip.yaml

# training with multiple gpus on Doppelgangers dataset with image flip augmentation
python train_multi_gpu.py doppelgangers/configs/training_configs/doppelgangers_classifier_flip.yaml

Citation

@inproceedings{cai2023doppelgangers,
  title     = {Doppelgangers: Learning to Disambiguate Images of Similar Structures},
  author    = {Cai, Ruojin and Tung, Joseph and Wang, Qianqian and Averbuch-Elor, Hadar and Hariharan, Bharath and Snavely, Noah},
  journal   = {ICCV},
  year      = {2023}
}