Detail of the student project

List
Topic:Aktivní učení pro sémantickou segmentaci mračen bodů
Department:Katedra kybernetiky
Supervisor:MSc. Ruslan Agishev
Announce as:Bakalářská práce, Semestrální projekt
Description:The aim of the project is the optimal selection of point cloud data samples for the training of semantic segmentation models in the sense of achieving greater accuracy with fewer training labels.

Method:
(a) Select model architecture that takes as input point cloud data and provides semantic labels for each point of the cloud. Proposal: The input point cloud data could be converted to range images. 2D CNN (Deeplab V3 [4]) could be used, which takes as input range images and produces semantic labels for each pixel.

(b) Select a data set to train the model. Proposal: SemanticKITTI [5] or SemanticUSL [6] (datasets have the same format).

(c) Train the model and compute performance metrics, Per-pixel Accuracy and mean Intersection over Union on the validation part of the data set.

(d) Implement an active data samples selection strategy from the same dataset to train the semantic segmentation model, which achieves a similar to baseline performance and requires fewer training samples. Proposal: Use the uncertainty sampling methods to query part of the data set (based on model confidence, entropy, query-by-committee, and margin sampling [1]).

(e) Train the model using the active learning strategy and report achieved performance and the required number of training samples. Compare to the baseline training method (selection of training samples randomly).

(f) Use localization data to provide consistent predictions for objects in the same environment observed from different positions.
Bibliography:[1] B. Settles, Active Learning Literature Survey, CS Technical Report, https://burrsettles.com/pub/settles.activelearning.pdf
[2] J. Prendki, An Introduction to Active Learning, ODSC, https://opendatascience.com/an-introduction-to-active-learning/
[3] Active Learning Tutorial, https://towardsdatascience.com/active-learning-tutorial-57c3398e34d
[4] L. Chen et al, Rethinking Atrous Convolution for Semantic Image Segmentation, https://arxiv.org/abs/1706.05587v3
[5] J. Behley et al, Towards 3D LiDAR-based semantic scene understanding of 3D point cloud sequences: The SemanticKITTI Dataset, http://semantic-kitti.org/
[6] P. Jiang et al, LiDARNet: A Boundary-Aware Domain Adaptation Model for Lidar Point Cloud Semantic, http://www.unmannedlab.org/research/SemanticUSL
Responsible person: Petr Pošík