In semi-supervised semantic segmentation, a model is trained with a limited number of labeled images to reduce the high annotation effort. While previous methods are able to learn good segmentation boundaries, they are prone to confuse classes with similar visual appearance due to the limited supervision. On the other hand, vision-language models (VLMs) are able to learn diverse semantic knowledge from image-caption datasets but produce noisy segmentation due to the image-level training. In SemiVL, we propose to integrate rich priors from VLM pre-training into semi-supervised semantic segmentation to learn better semantic decision boundaries. To adapt the VLM from global to local reasoning, we introduce a spatial fine-tuning strategy for label-efficient learning. Further, we design a language-guided decoder to jointly reason over vision and language. Finally, we propose to handle inherent ambiguities in class labels by providing the model with class definitions in text format. SemiVL achieves major gains over previous semi-supervised methods.

Latent diffusion models (LDMs) have demonstrated an extraordinary ability to generate creative content. However, are they usable as large-scale data generators, e.g., to improve tasks in the perception stack, like semantic segmentation? We investigate this question in the context of autonomous driving, and answer it with a resounding "yes". Our DGInStyle framework learns segmentation-conditioned image generation from synthetic data while preserving the original diverse style prior of LDMs. Using DGInStyle, we generate a diverse dataset of street scenes, train a semantic segmentation model on it, and evaluate the model on multiple popular autonomous driving datasets. DGInStyle improves the state of the art for domain-robust semantic segmentation by +2.5 mIoU.

Image-Text pretraining on web-scale image caption dataset has become the default recipe for open vocabulary learning thanks to the success of CLIP and its variants. However, the contrastive objective only focuses on image-text alignment and does not incentivise image feature learning for dense prediction tasks. In this work, we propose the simple addition of local-to-global correspondence learning by self-distillation as an additional objective for contrastive pre-training to propose SILC. Our model SILC sets a new state of the art for zero-shot classification, few shot classification, image and text retrieval, zero-shot segmentation, and open vocabulary segmentation.

In this paper, we extend our DAFormer (CVPR22) and HRDA (ECCV22) beyond synthetic-to-real domain adaptation to facilitate a more comprehensive view on domain robustness. In particular, we additionally show the capabilities of both for day-to-night and clear-to-adverse-weather domain adaptation. Further, we extend DAFormer and HRDA to domain generalization, enabling inference on unseen domains during test time. DAFormer and HRDA significantly improve the state-of-the-art domain adaptation and generalization performance by more than 10 mIoU on 5 different benchmarks.

In this work, we study domain adaptation for panoptic segmentation, which combines semantic and instance segmentation. As panoptic segmentation has been largely overlooked by the domain adaptation community, we revisit well-performing domain adaptation strategies from other fields, adapt them to panoptic segmentation, and show that they can effectively enhance panoptic domain adaptation. Further, we study the panoptic network design and propose a novel architecture (EDAPS) designed explicitly for domain-adaptive panoptic segmentation. EDAPS significantly improves the state-of-the-art performance for panoptic segmentation UDA by a large margin of 20% on SYNTHIA-to-Cityscapes and even 72% on SYNTHIA-to-Mapillary Vistas.

In unsupervised domain adaptation (UDA), a model trained on source data (e.g. synthetic) is adapted to target data (e.g. real-world) without access to target annotation. Most previous UDA methods struggle with classes that have a similar visual appearance on the target domain as no ground truth is available to learn the slight appearance differences. To address this problem, we propose a Masked Image Consistency (MIC) module to enhance UDA by learning spatial context relations of the target domain as additional clues for robust visual recognition. MIC enforces the consistency between predictions of masked target images, where random patches are withheld, and pseudo-labels that are generated based on the complete image by an exponential moving average teacher. To minimize the consistency loss, the network has to learn to infer the predictions of the masked regions from their context. Due to its simple and universal concept, MIC can be integrated into various UDA methods across different visual recognition tasks such as image classification, semantic segmentation, and object detection. MIC significantly improves the state-of-the-art performance across the different recognition tasks for synthetic-to-real, day-to-nighttime, and clear-to-adverse-weather UDA. For instance, MIC achieves an unprecedented UDA performance of 75.9 mIoU and 92.8% on GTA→Cityscapes and VisDA-2017, respectively, which corresponds to an improvement of +2.1 and +3.0 percent points over the previous state of the art.

Unsupervised domain adaptation (UDA) aims to adapt a model trained on synthetic data to real-world data without requiring expensive annotations of real-world images. As UDA methods for semantic segmentation are usually GPU memory intensive, most previous methods operate only on downscaled images. We question this design as low-resolution predictions often fail to preserve fine details. The alternative of training with random crops of high-resolution images alleviates this problem but falls short in capturing long-range, domain-robust context information. Therefore, we propose HRDA, a multi-resolution training approach for UDA, that combines the strengths of small high-resolution crops to preserve fine segmentation details and large low-resolution crops to capture long-range context dependencies with a learned scale attention, while maintaining a manageable GPU memory footprint. HRDA enables adapting small objects and preserving fine segmentation details. It significantly improves the state-of-the-art performance by 5.5 mIoU for GTA→Cityscapes and by 4.9 mIoU for Synthia→Cityscapes, resulting in an unprecedented performance of 73.8 and 65.8 mIoU, respectively.

As acquiring pixel-wise annotations of real-world images for semantic segmentation is a costly process, a model can instead be trained with more accessible synthetic data and adapted to real images without requiring their annotations. This process is studied in Unsupervised Domain Adaptation (UDA). In this work, we particularly study the influence of the network architecture on UDA performance and propose DAFormer, a Transformer network architecture tailored for UDA. It is enabled by three simple but crucial training strategies to stabilize the training and to avoid overfitting the source domain. DAFormer significantly improves the state-of-the-art performance by 10.8 mIoU for GTA→Cityscapes and by 5.4 mIoU for Synthia→Cityscapes.

We extend our CVPR21 paper "Three Ways to Improve Semantic Segmentation with Self-Supervised Depth Estimation" (see below) to semi-supervised domain adaptation featuring Cross-Domain DepthMix and Matching Geometry Sampling to align synthetic and real data.

Domain adaptation for semantic segmentation aims to improve the model performance in the presence of a distribution shift between source and target domain. In this work, we leverage the guidance from self-supervised depth estimation, available on both domains, to bridge the domain gap. On the one hand, we propose to explicitly learn the task feature correlation to strengthen the target semantic predictions with the help of target depth estimation. On the other hand, we use the depth prediction discrepancy from source and target depth decoders to approximate the pixel-wise adaptation difficulty. The adaptation difficulty, inferred from depth, is then used to refine the target semantic segmentation pseudo-labels.

We developed a method for improving semantic segmentation based on knowledge learned by self-supervised monocular depth estimation from unlabelled image sequences. In particular, (1) we transferred knowledge from features learned during self-supervised depth estimation to semantic segmentation, (2) we implemented a strong data augmentation by blending images and labels using the structure of the scene, and (3) we utilized the depth feature diversity as well as the level of difficulty of learning depth in a student-teacher framework to select the most useful samples to be annotated for semantic segmentation.

We extended saliency maps from classification to dense predictions to allow visual inspection of semantic segmentation convolutional neural networks. We investigated the effectiveness of the proposed Grid Saliency on a synthetic dataset with an artificially induced bias between objects and their context as well as on real-world datasets. Our results show that Grid Saliency can be successfully used to provide easily interpretable context explanations and, moreover, can be employed for detecting and localizing contextual biases present in the data.

We developed a self-supervised temporal prediction and multi-camera fusion system based on agent-centric semantic maps. Semantic information from multiple cameras is integrated over multiple frames in a unified semantic bird’s eye view environment representation.

We designed and evaluated an external, camera-based localization and identification system for swarm robots using convolutional neural networks. For a convenient system setup, we developed a low-effort training data acquisition and synthetization process.

In the RoboCup @work team "RobOTTO", I helped to build an autonomous mobile robot for the transport of work items in factories. From 2015 until 2018, I was responsible for the development of the state machine, world model, task planner, and object recognition. In 2017, we achieved the 2nd place at the World Cup.

For the RoboCup Junior league Rescue-B (now Rescue Maze), we built and programmed a robot from scratch to autonomously search for heat sources in a maze, which simulate victims in a building in danger of collapsing. In 2013 and 2014, we achieved the 1st place at the World Cup competitions.

For the "Jugend forscht" German high school research competition, we developed a cost-effective optical spectrometer for schools, which is more than 90% cheaper than regular devices. At the federal stage, we received the award for non-destructive testing. After the competition, I scaled the prototype to small batch production and equipped about 30 schools with the spectrometer.