Skip to content

Research links

Eirik edited this page May 31, 2019 · 6 revisions

Dataset:

Any relevant dataset, can compare the type of data, dimensions, data processing methods and how it has been collected/with what equipment.

ScanNet

Description: Dataset containing 2D (+Depth) and 3D data. Really large (2D RGB-D contains ≈2.5 million frames).

Pascal2 / PASCAL Visual Object Classes / VOC

NYU Depth Dataset V2

Stanford 2D-3D-Semantics Dataset

SUN RGB-D: A RGB-D Scene Understanding Benchmark Suite

Data Processing:

External files/documents that could be relevant when processing my data

RGB-D Semantic Segmentation:

Descriptive webpages:

Depth-aware CNN for RGB-D Segmentation (2018):

Description: State of the art (?) research on RGB-D, use one of the models as baseline. They have models both for depth images and for HHA encoded depth data.

Learning Rich Features from RGB-D Images forObject Detection and Segmentation (2014):

Comment: Introducing (?) RGB-D and HHA for semseg

Cross Modal Distillation for Supervision Transfer (2015)

Comment: 'Paired' networks, 'continuation' from gupta2014learning

Depth Not Needed - An Evaluation of RGB-D Feature Encodings for Off-Road Scene Understanding by Convolutional Neural Network (2018):

Comment: Discussing the 'efficiency' of depth data (and different types of depth data), Paper conclusion:

Our results show that the additional information contained in stereo disparity can provide marginal improvement to CNN classification performance compared to standard RGB images in this particularly challenging problem of off-road scene understanding. However, careful consideration of how this information is encoded is necessary, and in our case the combination of RGB data with each pixel’s height above ground plane was shown to give the best results. The performance gains demonstrated are small, however this is mostly due to the efficacy of the existing CNN architecture at classification from conventional colour images (RGB), with results so good to begin with that it becomes ever harder to gain extra performance. This is counter to results such as those from Gupta et al [8] and Pavel et al [9], who claim that RGBD data provides significant improvements in CNN classification performance over RGB imagery. This demonstrates that while depth information may improve CNN performance in an indoor object detection task, it is of limited utility in the more challenging off-road environment, where poorly defined class boundaries and inconsistent object shape can hamper classification accuracy.

ENetDepth

Description: Fast net for RGB-D segmentation tasks Original name: ENet: A Deep Neural Network Architecture forReal-Time Semantic Segmentation The original was made for RGB (?)

AdapNet++ / SSMA

Description: Researchers from the university in Freiburg! New (Jan 2019)

FuseNet: Incorporating Depth into Semantic Segmentation via Fusion-Based CNN Architecture (2018)

RGB-D Semantic segmentation with Fusion-based CNN Architecture

RGB - Point Cloud

Minowski Engine / 4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks

Description: State of the art engine and models for segmentation using rgb point clouds.

Paper GitHub