Stellenangebote

Abschlussarbeiten

Masterarbeit: Video Prediction in Label Space

Example of intended results. Input sequence of label images from the Cityscapes dataset (left) and desired output sequence (right).

Description:
State of the art machine learning models for video generation have shown, that predicting future frames of low dimensional synthetic videos works quite good. However, predicting natural videos in high quality still leaves room for improvement due to the high complexity of natural scenes.

Your Master's Thesis will focus on extending one of the existing models for video generation and prediction. The resulting method should be able to predict a sequence of probable future label images given the label images of an input video sequence. For this task, you are free to choose which of the existing methods you want to use. One difficulty of adapting an existing model will be to prevent the model from generating blurry versions of the input images. The results of your final model have to be evaluated both quantitatively and qualitatively, and have to be compared to results of related approaches.

For the purpose of this topic, it is of advantage to have practical experience with deep learning frameworks and a good understanding of machine learning principles. Also, good programming skills in Python are essential. All further details will be settled in person.

Requirements:

  • Completed course(s) in machine learning or equivalent
  • Advanced programming skills (Python)
  • Practical experience with deep learning frameworks (e.g. TensorFlow, Torch, Caffe, etc.)

Contact:

Masterarbeit: Conditional Synthesizing of Photo-Realistic Images from Label Images

Exemplary results of Chen and Koltun's Cascaded Refinement Network. Label image of the Cityscapes dataset (left) and corresponding synthesized image (right). [Photographic Image Synthesis with Cascaded Refinement Networks, Chen and Koltun, ICCV 2017]

Description:
State of the art machine learning models for image generation have shown, that synthesizing images given their corresponding label image yields to good results, even for complex natural scenes, such as images of traffic scenes. Since the training data of these models consist of pairs of label images and corresponding real world images, the learned features for each label class represent only an average over all examples. Hence, those models will fail to display details of individual scenes, e.g., color information, correctly.

Your Master's Thesis will focus on extending one of these models. The main goal of your work will be to implement an online fine-tuning mechanism that allows the model to synthesize details of a given scene correctly. In the end, the synthesized image will be based on the label information and also be conditioned on information from the original scene. The results of your final model have to be evaluated both quantitatively and qualitatively, and have to be compared to results of related approaches.

For the purpose of this topic, it is of advantage to have practical experience with deep learning frameworks and a good understanding of machine learning principles. Also, good programming skills in Python are essential. All further details will be settled in person.

Requirements:

  • Completed course(s) in machine learning or equivalent
  • Advanced programming skills (Python)
  • Practical experience with deep learning frameworks (e.g. TensorFlow, Torch, Caffe, etc.)

Contact:

Studentische Hilfskräfte

Setting up a Simulation Framework for Autonomous Driving