3 ms·
This is not "modern" though, the paper is quite old. He has improved it significantly since then: https://arxiv.org/abs/1612.08242 https://arxiv.org/abs/1612.08
by 0xbear 9y ago
This is not "modern" though, the paper is quite old. He has improved it significantly since then: https://arxiv.org/abs/1612.08242 https://arxiv.org/abs/1612.08242. Best paper honorable mention at CVPR2017.
- hedgehog 9y agoI missed that the site was updated for YOLO9000, it is very impressive. The Google paper is recent though (July), maybe you're thinking of https://arxiv.org/abs/1611.10012 https://arxiv.org/abs/1611.10012? Anyway, what I meant to point out was that there has been a lot of work since on improving the efficiency of the architectures used for feature extraction (Xception, MobileNet, ShuffleNet, NASNet).
- 0xbear 9y agoSure, there's new work all the time. But I think you might be missing the forest for the trees here. The main efficiency gain here is not in how you extract and transform features (although those improvements would also work in YOLO, as far as I can tell). The novel part is in structuring the problem as a single forward pass regression. Further improvement in this year's paper is in a better loss function. Other possible efficiency improvements are orthogonal and complementary to that.
- hedgehog 9y agoI agree that the single shot approach in YOLO was pretty clearly a big step forward. In the two years since though there much of the efficiency work has been in the underlying feature detector architecture, which as you point out should integrate well with the YOLO9000 training improvements. It's very cool to see this kind of capability get within range of phone-sized devices.