2D Scene Parsing Dataset of TUM CMS Indoor Point Cloud

2D Scene Parsing Dataset of TUM CMS Indoor Point Clouds is generated as a benchmarking datasets in 2D understanding of existing TUM CMS Indoor Point Cloud, available on mediaTUM. The dataset includes 219 raw images that are taken from indoor environments of TUM, with different sizes, as *.jpg files provided by the responsible authors of the TUM CMS Indoor Point Cloud dataset. Raw images are provided either as images or 2D captures from TUM CMS Indoor Point Cloud. The dataset also includes 2D semantic annotations and depth information of raw images. 2D semantic segmentation annotations are derived from 150-class colored RGB format of ADE20K dataset, and presented as *.jpg files with the same sizes of raw images. The depth information of each raw image is provided as greyscale *.png files, including pixel values between 0-255, with the same sizes of raw images.      «

2D Scene Parsing Dataset of TUM CMS Indoor Point Clouds is generated as a benchmarking datasets in 2D understanding of existing TUM CMS Indoor Point Cloud, available on mediaTUM. The dataset includes 219 raw images that are taken from indoor environments of TUM, with different sizes, as *.jpg files provided by the responsible authors of the TUM CMS Indoor Point Cloud dataset. Raw images are provided either as images or 2D captures from TUM CMS Indoor Point Cloud. The dataset also includes 2D sem...     »

2D segmentation and depth prediction tasks are conducted on 219 raw images. For 2D segmentation, ViT-L+Mask2Former baseline architectures, pre-trained on the ADE20K dataset together with methods applied in SERNet-Former, are applied. The segmentation tasks are completed in 3 minutes 59 seconds in PyTorch virtual environment using NVIDIA A100 GPU hardware. Depth information generation is completed via ViT-L+Mask2Former baseline architecture in 44 seconds in PyTorch virtual environment using NVIDIA A100 GPU.     «

2D segmentation and depth prediction tasks are conducted on 219 raw images. For 2D segmentation, ViT-L+Mask2Former baseline architectures, pre-trained on the ADE20K dataset together with methods applied in SERNet-Former, are applied. The segmentation tasks are completed in 3 minutes 59 seconds in PyTorch virtual environment using NVIDIA A100 GPU hardware. Depth information generation is completed via ViT-L+Mask2Former baseline architecture in 44 seconds in PyTorch virtual environment using NVIDI...     »