EN FR
EN FR
RITS - 2018
New Software and Platforms
Bilateral Contracts and Grants with Industry
Bibliography
New Software and Platforms
Bilateral Contracts and Grants with Industry
Bibliography


Section: New Results

Convolutional neural networks for Semantic and Completion with Sparse and Dense Data

Participants : Raoul de Charette, Maximilian Jaritz, Fawzi Nashashibi.

Deep convolutional networks have outperform all previous techniques on most vision tasks. This is because they are able to utilize dense data and extract relationship between local information such as gradients, or high level features. However, convolutional neural networks (CNNs) require dense data and are known to fail when data is sparse. Here, we address the research problem and proposed a solution. Instead of using a sparse convolution methodology, we show that using the right architecture with a proper training strategy the network can learn sparsity invariant feature while remaining stable when dense data are present. Our architecture uses an encoder-decoder version of Mobile NasNet with skip connections. The results show that we can accomplish both data completion or semantic segmentation changing only the last layer of the network. Performance obtained were published on Kitti Benchmark and ranks among the first ones, and the methodology was published in 3DV [26]. This research was partially funded by Valeo.