TLDR
Sparse R-CNN is a new method that works with sparse convolutions on 3D computer vision tasks such that. It achieves near state-of-the-art performance in object detection and uses completely sparse and learnable bounding boxes generation. The architecture is elegant. It consists of FPN based backbone that acquires features from images, and. extractable features from. and. the main contribution to neural nets architecture of this very very. interactive head. Each feature is fed into its own “exclusive feature for object and classification, where each head is conditioned on specific. feature”via the TL;DR App
no story
Written by zetyquickly | Machine learning enthusiast. Research engineer at Skoltech.