Posted on

Football Playing Field Registration Using Distance Maps


In football, the playing field registration problem can be thought of as the homography estimation between the playing field plane and the visible area of
the image. Existing approaches rely on the identification of significant keypoints in the image, such as the corners or intersections of the playing
field lines, which are then used for the initial estimate of the homography. Finding the exact location of such keypoints can be very challenging, and
that is why the initial homography estimates are not accurate enough and need to be refined iteratively. We propose distance maps, a technique that
utilizes distance transform to predict the location of imaginary lines spread evenly over the playing field. Intersections of these lines can be
used to find more accurate homography estimates. We also introduce the Calib360 playing field registration dataset, that can provide a sufficient amount of
training data with accurate ground truth labels.


The source code of our method can be found at – … URL …

The best pre-trained model can be downloaded at – … URL …

You can find our original paper at – … URL …


We propose the Calib360 dataset, which contains 300,000 training images with accurate homography annotations, which we have constructed from equirectangular panoramas. In panoramas, we annotate the positions of known keypoints of the football playing field model.

Next, from the annotated keypoints we extract the panoramic camera position and orientation with respect to the playing field. Once the camera parameters are extracted, we can generate unlimited number of different views of the playing field by adjusting the viewing angles and field of view. These generated images are perfect for training of neural networks because they all conform to the pinhole camera model and contain accurate annotations.


Estimated playing field registration by our method (first column), by the method of Nie et al. (2021, second column), by the method of Chu et al. (2022, third column), and by the method of Theiner et al. (2023, fourth column).






Posted on

Improving Radial Lens Distortion Correction Using Multi-task Learning


Sports image analysis is one of the domains that has significantly benefited from advances in computer vision. With the growing adoption of machine learning methods new products and tools have emerged that make sports image analysis both accurate and accessible for professional athletes, amateur and junior athletes, and fans all around the world. Sports images serve as crucial visual records, aiding in post-game or live analysis and decision-making. However, very often they are affected by radial distortion that hinders their accurate interpretation and utilization.

We propose a deep-learning regression-based method to rectify images containing radial lens distortion, that operates on single independent images. Our method hase been trained on the football domain dataset, and has been fine-tuned specifically for images containing small radial distortion, which is more difficult to rectify accurately. Compared to other deep-learning radial distortion correction methods, we introduce the following contributions:

  • Secondary learning task that learns useful distortion features on random combinations of sample pairs, prevents overfitting, and encourages the feature extractor to learn more general distortion features.
  • Penalty term that encourages better accuracy on low-distortion images

The official GitHub repository for our method can be found at –

The best trained model and training configuration can be downloaded here – improving-radial-best.tar.gz

The original paper can be found here –


We have performed a quantitative evaluation of our proposed method on the validation set of the Football360 validation subset, and compared our method with 3 other methods. Our method outperformed all other evaluated methods on all metrics by a significant margin.

Janos and Benesova (2023)0.85223.075 dB0.0172
Li et al. (2019)0.41710.919 dB1.431
Santana-Cendres et al. (2016)0.71421.647 dB0.1132
Our method0.93330.098 dB0.0088

We have performed a qualitative evaluation of our method on the World Cup 2014 dataset. The WorldCup 2014 images have not been used during the training of neither of the evaluated methods.

Input imagesCorrected by our methodJanos and Benesova (2023)Li (2019)Santana-Cedres (2016)




Posted on

Football360 – Introducing a New Dataset for Camera Calibration in Sports Domain


The aim of this dataset is to aid the development and evaluation of computer vision algorithms, primarily the radial distortion correction algorithms, in the sports domain.

This dataset contains 268 panorama images, and was created using the PANONO panoramic camera in 3 football arenas in Slovakia. Each arena was covered from numerous locations on all levels of the tribune, and broadcast camera platforms. The images capture regular football game, pitch maintenance, low/challenging lighting conditions, day and night situations.

ArenaNumber of images
Dunajská Streda53

You can download the raw images from the following link – football360-raw.tar.gz (7.7 GB)

There is also an official GitHub repository available at –

The original paper can be found here –

Raw Images

Raw images are stored as 16384×8192 JPG, they are the direct result of the PANONO stitching service.


Exported Sets

There are four exported datasets available for direct download and use. Each dataset contains a collection of images (data) distorted with known distortion coefficients (labels).

Set namePurposeImagesPresetDownload Link
ATraining30,000setA.jsonfootball360-setA.h5 (10.5 GB)
BTraining100,000setB.jsonfootball360-setB.h5 (35.2 GB)
CTraining300,000setC.jsonfootball360-setC.h5 (105.5 GB)
VValidation10,000setV.jsonfootball360-setV.h5 (3.5 GB)

You can see an example of how to load and use these files in a linked notebook –

Distorted ImageCorrected Image


Igor Janos

Posted on

Explainable 3D convolutional neural network using GMM encoding

Abstract:     The aim of this paper is to propose a novel method to explain, interpret, and support the decision-making process of deep Convolutional Neural Network (CNN). This is achieved by analyzing neuron activations of trained 3D-CNN on selected layers via Gaussian Mixture Model (GMM)   …  more


Posted on

Automatic Music Transcription using WaveNet

Deep generative models such as WaveNet are surprisingly good at various modelling tasks. We exploit the modelling capacity of WaveNet architecture in a setup that is quite different from the original generative case: for feature extraction and pattern recognition in sake of polyphonic music transcription. The model is trained end-to-end to perform the underlying task of multiple fundamental frequency estimation by processing raw waveforms of digital audio signal.   …  more