Daily CVPR - Monday

Tanushri: The main concept of my idea within my tracking algorithm is probability. So here we are trying to estimate the probability of an object being present in a video. So let’s say I used some features specific to the object. Let’s say that if you want to track a face, I want to track something like the nose or the lips, I will use those features, and those features will try to estimate a probability that yes, it is present in future frames as well. CVPR Daily: Can you tell me what was challenging in this work? Tanushri: The most challenging aspect, I would say, is to find robust features of an object, because if an object is there, but you are unable to find robust features, or you find features which are constantly changing. So for us, as humans, we can still differentiate between the same type of objects having different poses or they change differently, like there is dark thing and he goes there, and you can say yes, it is that [same] person. But for a computer, everything becomes the same because it is still very nascent and very naive. So you need to give the computer version algorithm something that is strong enough to tell that these are the features specific to that object. CVPR Daily: What are the next steps for this work? Tanushri: That next step are… since I’m the fourth in the tracking benchmark, so my objective is to make it the top performing algorithm, not the fourth. So I want to combine a detector also to my tracking algorithm so that I can have better precision and success rates. CVPR Daily: Can you tell me something unexpected that happened? Maybe you have some funny story to tell us. What did you not expect to happen... and it did happen. Tanushri: Yea, ok… so the funny or the challenges that you really couldn’t think about before you started testing on data was something that you are visually seeing in the video. Let’s say you see a person coming in the video. Then suddenly, he or she disappears. So for a human, it’s like “ok it will come again or maybe it won’t come again”, but for the tracker it moves here and there in no sequence constantly. Like it’s saying that, ok, that wall is also an object. So it’s very funny to see how stupid an algorithm can be. CVPR Daily: Monday Presentation 7 “ It’s very funny to see how stupid an algorithm can be ”