Holistic Multi-View Building Analysis in the Wild with Projection Pooling
Authors: Zbigniew Wojna, Krzysztof Maziarz, Łukasz Jocz, Robert Pałuba, Robert Kozikowski, Iason Kokkinos2870-2878
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We address six different classification tasks related to finegrained building attributes... We introduce a new benchmarking dataset... We propose a new projection pooling layer... Introducing this layer improves classification accuracy compared to highly tuned baseline models indicating its suitability for building analysis. |
| Researcher Affiliation | Collaboration | 1 Tensorflight 2 Jagiellonian University 3 University College London; *Work done during an internship at Tensorflight. The author is now at Microsoft Research. |
| Pseudocode | Yes | Algorithm 1 Pool(h , w , prt, si, srci, diri, Fo Vi) |
| Open Source Code | No | The paper does not provide an explicit statement or a link to a repository for the open-source code of their methodology. |
| Open Datasets | No | We introduce a new benchmarking dataset, consisting of 49426 images (topview and street-view) of 9674 buildings. Our dataset consists of 6477 training and 3197 testing building scenes, split by stratified sampling. The paper describes a newly created dataset but does not provide concrete access information (e.g., a link or DOI) for public download. |
| Dataset Splits | No | Our dataset consists of 6477 training and 3197 testing building scenes, split by stratified sampling. The paper explicitly states training and testing splits, but does not provide specific details (e.g., percentages or counts) for a separate validation split, even though hyperparameter tuning is mentioned. |
| Hardware Specification | Yes | All of our experiments were performed on a single GPU (NVIDIA Tesla V100 or RTX 2080 Ti) with 11GB of VRAM. |
| Software Dependencies | No | As the feature extractor network, we use the Res Net-50 model pre-trained on the Image Net dataset, as available in Py Torch (Paszke et al. 2017). The paper mentions PyTorch but does not provide specific version numbers for software dependencies. |
| Experiment Setup | Yes | We train the network using stochastic gradient descent with momentum... with a batch size of one building... We use the learning rate of 0.0001 for ten epochs and then 0.00001 for one more epoch. We set the momentum to 0.9. We apply L2 regularization with a weight decay of 0.001 and augment the dataset with random color jittering. |