MIT researchers develop new algorithm for improving object recognition

Researchers at MIT have developed a new algorithm that could improve and speed up object recognition. The code is particularly useful for robots that are used, for example, to perform simple household tasks.

When recognizing objects, the chance of correct recognition by an algorithm is significantly increased if an image sensor shoots images from multiple angles instead of just one viewing angle. Based on this fact, MIT researchers have adapted an existing algorithm that is used in radar systems, among other things. In addition, the algorithm generates multiple ideas about what an object might be by comparing this hypothesis with other images, MIT writes.

In order not to let the recognition mechanism become too slow, only the most probable hypothesis is retained, while the less probable recognitions are removed. Despite this, the algorithm is still quite slow because the most recently established hypothesis has yet to be verified.

The MIT researchers came up with the idea to modify the algorithm in such a way that not all assumptions are removed, but only a part. Samples are randomly selected from the total for selection. Because on average there is a lot of overlap in the hypotheses, consensus will eventually arise about what the object is very likely, with a certain number of samples. Due to a number of clever comparison methods with images from different angles, the algorithm would be considerably faster than the ‘old’ method.

According to the researchers, the algorithm can be used well for robots that need to be able to recognize objects quickly. During tests, the main focus was on objects that can be found in a household.