/surface_normal_uncertainty

[ICCV 2021 Oral] Estimating and Exploiting the Aleatoric Uncertainty in Surface Normal Estimation

Primary LanguagePythonMIT LicenseMIT

Estimating and Exploiting the Aleatoric Uncertainty in Surface Normal Estimation

PWC PWC

Official implementation of the paper

Estimating and Exploiting the Aleatoric Uncertainty in Surface Normal Estimation
ICCV 2021 [oral]
Gwangbin Bae, Ignas Budvytis, and Roberto Cipolla
[arXiv] [youtube]

The proposed method estimates the per-pixel surface normal probability distribution, from which the expected angular error can be inferred to quantify the aleatoric uncertainty. We also introduce a novel decoder framework where pixel-wise MLPs are trained on a subset of pixels selected based on the uncertainty. Such uncertainty-guided sampling prevents the bias in training towards large planar surfaces, thereby improving the level of the detail in the prediction.

Getting Started

We recommend using a virtual environment.

python3.6 -m venv --system-site-packages ./venv
source ./venv/bin/activate

Install the necessary dependencies by

python3.6 -m pip install -r requirements.txt

Download the pre-trained model weights and sample images.

python download.py && cd examples && unzip examples.zip && cd ..

[25 Apr 2022] download.py does not work anymore. Please download the models and example images directly from this link, and unzip them under ./checkpoints/ and ./examples/.

Running the above will download

  • ./checkpoints/nyu.pt (model trained on NYUv2)
  • ./checkpoints/scannet.pt (model trained on ScanNet)
  • ./examples/*.png (sample images)

Run Demo

To test on your own images, please add them under ./examples/. The images should be in .png or .jpg.

Test using the network trained on NYUv2. We used the ground truth and data split provided by GeoNet.

Please note that the ground truth for NYUv2 is only defined for the center crop of image. The prediction is therefore not accurate outside the center. When testing on your own images, we recommend using the network trained on ScanNet.

python test.py --pretrained nyu --architecture GN

Test using the network trained on ScanNet. We used the ground truth and data split provided by FrameNet.

python test.py --pretrained scannet --architecture BN

Running the above will save the predicted surface normal and uncertainty under ./examples/results/. If successful, you will obtain images like below.

The predictions in the figure above are obtained by the network trained only on ScanNet. The network generalizes well to objects unseen during training (e.g., humans, cars, animals). The last row shows interesting examples where the input image only contains edges.

Training

Step 1. Download dataset

  • NYUv2 (official): The official train/test split contains 795/654 images. The dataset can be downloaded from this link. Unzip the file nyu_dataset.zip under ./datasets, so that ./datasets/nyu/train and ./datasets/nyu/test/ exist.

  • NYUv2 (big): Please visit GeoNet to download a larger training set consisting of 30907 images. This is the training set used to train our model.

  • ScanNet: Please visit FrameNet to download ScanNet with ground truth surface normals.

Step 2. Train

  • If you wish to train on NYUv2 official split, simply run
python train.py
  • If you wish to train your own model, modify the file ./models/baseline.py and add --use_baseline flag. The default loss function UG_NLL_ours assumes uncertainty-guided sampling, so this should be changed to something else (e.g. try --loss_fn NLL_ours).

Citation

If you find our work useful in your research please consider citing our paper:

@InProceedings{Bae2021,
    title   = {Estimating and Exploiting the Aleatoric Uncertainty in Surface Normal Estimation}
    author  = {Gwangbin Bae and Ignas Budvytis and Roberto Cipolla},
    booktitle = {International Conference on Computer Vision (ICCV)},
    year = {2021}                         
}