Segmentation de scènes extérieures à partir d'ensembles d'étiquettes à granularité et sémantique variables

Abstract : In this work, we present an approach that leverages multiple datasets annotated using different classes (different labelsets) to improve the classification accuracy on each individual dataset. We focus on semantic full scene labeling of outdoor scenes. To achieve our goal, we use the KITTI dataset as it illustrates very well the focus of our paper : it has been sparsely labeled by multiple research groups over the past few years but the semantics and the granularity of the labels differ from one set to another. We propose a method to train deep convolutional networks using multiple datasets with potentially inconsistent labelsets and a selective loss function to train it with all the available labeled data while being reliant to inconsistent labelings. Experiments done on all the KITTI dataset's labeled subsets show that our approach consistently improves the classification accuracy by exploiting the correlations across data-sets both at the feature level and at the label level.
Document type :
Conference papers
Complete list of metadatas

Cited literature [33 references]  Display  Hide  Download

https://hal.archives-ouvertes.fr/hal-01318461
Contributor : Damien Fourure <>
Submitted on : Friday, May 20, 2016 - 4:21:58 PM
Last modification on : Monday, January 13, 2020 - 5:46:06 PM
Long-term archiving on: Sunday, August 21, 2016 - 10:23:56 AM

File

Segmentation_scenes_exterieure...
Files produced by the author(s)

Identifiers

  • HAL Id : hal-01318461, version 1

Citation

Damien Fourure, Rémi Emonet, Elisa Fromont, Damien Muselet, Alain Tremeau, et al.. Segmentation de scènes extérieures à partir d'ensembles d'étiquettes à granularité et sémantique variables. RFIA 2016, Jun 2016, Clermont Ferrand, France. ⟨hal-01318461⟩

Share

Metrics

Record views

368

Files downloads

578