Details

Type

Bachelor Thesis
Student Project
Master Thesis

Persons

1-3

Motivation

We want to classify objects in scenes that contain many similar objects (e.g., in a factory hall, warehouse, or office). However, especially for non-standard objects, training these specific classes requires a lot of effort to create enough ground truth. Instead, we aim at classifying yet unseen objects automatically and label them just with weak supervision, i.e., a human-in-the-loop being queried for unknown classes for training on-the-fly.

Description

Unsupervised object detection [1] can also be done using open vocabulary object detection [2], or zero-shot learning [3]. Objects can be divided into classes by k-means clustering, see this BA for a simple 2-means image classification [4]. Here is a survey in weakly supervised object detection [5].

[1] https://openaccess.thecvf.com/content/CVPR2021/papers/Tian_Unsupervised_Object_Detection_With_LIDAR_Clues_CVPR_2021_paper.pdf

[2] https://arxiv.org/abs/2207.01987

[3] https://arxiv.org/abs/2104.04980

[4] https://www.cg.tuwien.ac.at/research/publications/2019/Gruber2019/Gruber2019-Bachelor%20Thesis.pdf

[5] https://arxiv.org/pdf/2104.07918.pdf

Tasks (extent depends on PR/BA/DA)

  • Segment objects: apply a suitable ML algorithm
  • Cluster similar objects: apply a suiting ML algorithm
  • Label objects in clusters with human-in-the-loop and adjust clusters

 

Requirements

Python and C++ programming skills and interest in geometry processing and machine learning.

Environment

Platform-independent.

A bonus of €500/1,000 if completed to satisfaction within an agreed time frame of 6/12 months (PR/BA or DA).

Responsible

For more information please contact Stefan Ohrhallinger, Michael Wimmer.