Home
Uni-Logo
 

Oberseminar Bildverarbeitung, Computersehen und Computergraphik

Prof. Dr.-Ing. Thomas Brox, Prof. Dr.-Ing Matthias Teschner, apl. Prof. Dr. Olaf Ronneberger

Nächste Termine:
10.01.2019
11.01.2019

Vorankündigung (Aushang):

Donnerstag, 10.01.2019, 13.00 hst
Geb. 52, SR 01-033

Herr Jan Bechtold
Institut für Informatik, Universität Freiburg

berichtet über die Ergebnisse seiner Masterarbeit:

3D Object Detection using Tangent Convolutions



Object detection is one of the classic computer vision tasks. It requires localizing individual objects in visual data and determining their type. State of the art methods for 2D object detection use convolutional neural networks (CNNs), which can be efficiently applied to a dense grid of information, like images. In contrast to images, applying CNNs to 3D data is not straightforward because 3D grid convolution is inefficient for large scenes.

In this work we build an object detection framework for 3D point clouds. It is based on Tangent Convolutions and follows the Faster R-CNN architecture. Tangent Convolutions efficiently implement a CNN for analyzing 3D point clouds, by convolving local approximations of a 3D structure with a 2D kernel.

Our object detector is efficient and scales to large scenes with hundreds of thousands of points, due to the use of tangent convolutions. We evaluate our method on two indoor datasets: ScanNet and Stanford Large-Scale 3D Indoor Spaces Dataset (S3DIS). The proposed method is generic and can be applied to both indoor datasets without any changes. Experimental results show that our object detector is able to detect objects of various sizes and shapes, ranging from a small sink to a large couch. A comparison of our ScanNet benchmark score to other groups on the leader board shows that our method is competitive with other approaches.



Bitte beachten Sie die Zeit und den Ort!

Interessenten sind herzlich eingeladen. Weitere Informationen bei: Prof. Dr.-Ing. Thomas Brox, Tel: 0761/203-8261


Vorankündigung (Aushang):

Freitag, 11.01.2019, 15.00 hst
Geb. 52, SR 01-033

Herr Kangkan Bharadwaj
Institut für Informatik, Universität Freiburg

berichtet über die Ergebnisse seiner Masterarbeit:

Learning depth-based semantic segmentation of street scenes



This work addresses multi-class semantic segmentation of street scenes by ex-ploring depth information with RGB data. Our dataset comprises of street images from Berlin taken from four different camera angles and scanned using a laser scanner and later processed to create the depth images from 3D point clouds by projection. Our work also proposes an architecture model comprising of a Residual Network as an encoder and a UNet decoder for the Berlin set that learns good quality feature representation. We achieve a mean accuracy of 58.35%, mean pixel accuracy of 94.36% and mean IOU (Intersection over Union) of 51.91% on the test set. We further analyze the benefits that the model exhibits on certain classes when trained including depth to the RGB data with that of the model based only on RGB information. An alternative approach of feeding the depth information using a separate encoder was carried out to study the performance variation in segmentation and if it can bring any significant hike to its quality. And finally we draw a performance contrast of our network to one of the state-of-the-art models on our dataset.



Bitte beachten Sie die Zeit und den Ort!

Interessenten sind herzlich eingeladen. Weitere Informationen bei: Prof. Dr.-Ing. Thomas Brox, Tel: 0761/203-8261