What is sponge ?

Sponge is an open source project trying to solve a specific problem: training and evaluating objects detection algorithms.

This software aims to make the life of researchers and new students easier and better.

Why ?

Having people rewrite the same program over and over is evil.
Creating an opensource community to solve a common, well identified problem, is good.

Today there is a large community doing research on image processing. Many many researchers do classification, detection and recognition of images. Yet everytime they have to write from scratch:

• The dataset parsing code.
• The feature descriptor computation and/or the machine learning algorithm.
• The glue between parsing the data, computing the features, training the machine learning algorithm, testing the algortihm and evaluating the algorithm.
• The plotting code to compare the results.

It is time to stop the madness. We should have an open source tool that provides all the common bits, the glue and allows us to focus on what we really care: the algorithms.

The dream

I would like to have a one-stop-shop to access the state of the art images classifiers, detectors and recognizers.

Goal

The goal can be described in three user stories:

New user

1. New user discovers the tool. On his ubuntu machine he should be able to copy paste a set of ~5 command lines and after ~20 minutes he got a trained, usable, pedestrian detector, from which he obtained a "miss rate versus false positives per image" graph that compares the new detector with existing state of the art.

New detector
2. User wants to integrate its own classifier, or feature descriptor. The user should be able to visit a well identified website, click on a big "tutorials" button, follow step by step instructions and be able to download an example project.

New dataset
3. A laboratory has created a new dataset. He wants to make it publicly available in a format easy to use for other laboratories. The user should be able to visit a well identified website and post there a link so other users can use the dataset directly (in a way similar to ruby gems).