Fermi Denoising: Evaluation
The problem is an object detection problem. Each sample (an observation) is np.array of 200x200, presenting number of photons per pixel. You must predict the positions of point sources (we are not interested in the brigthness for the moment).
You are given for training X_train images and y_train array of corresponding source coordinates. You must train a model which predicts an array of point sources coordinates for a given image.
There are 2 phases:
- Phase 1: development phase. We provide you with training data and its corresponding list of point source coordinates and validation and test data without list of corrdinates. Make predictions for both datasets. However, you will receive feed-back on your performance on the validation set only. The performance of your LAST submission will be displayed on the leaderboard.
- Phase 2: final phase. You do not need to do anything. Your last submission of phase 1 will be automatically forwarded. Your performance on the test set will appear on the leaderboard when the organizers finish checking the submissions.
This competition allows you to submit either:
- A pre-trained prediction model.
- A prediction model that must be trained and tested.
The submissions are evaluated using the chosen metric. This metric computes the sum of the distances from each real source to the closest source from the predicted list of sources and the distances from each predicted sources to the closest source from the list of real sources. The perfect prediction returns 0.
Submissions must be made before the end of phase 1.
Fermi Denoising: Data
This challenge relies on the dataset generated from fermi tools and simulating the real fermi images.
Each file from input_data is a numpy array 200x200, an image which contains an observation.
Each file from output_data is an array nx2, where n is the number of point sources. It conntains coordinates corresponding to pixels of the image. ()x and y from 0 to 200)