Learning to Downsample for Segmentation of Ultra-High Resolution Images

Authors: Chen Jin, Ryutaro Tanno, Thomy Mertzanidou, Eleftheria Panagiotaki, Daniel C. Alexander

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on benchmarks of high-resolution street view, aerial and medical images demonstrate substantial improvements in terms of efficiency-and-accuracy trade-off compared to both uniform downsampling and two recent advanced downsampling techniques.
Researcher Affiliation Collaboration 1 Centre for Medical Image Computing, Department of Computer Science, University College London 2 Healthcare Intelligence, Microsoft Research Cambridge {chen.jin,t.mertzanidou,e.panagiotaki,d.alexander}@ucl.ac.uk rytanno@microsoft.com
Pseudocode No The paper does not contain a clearly labeled "Pseudocode" or "Algorithm" block, nor structured steps formatted like pseudocode.
Open Source Code Yes Video demos available at https://lxasqjc.github.io/learn-downsample.github.io/. This site contains a link to the GitHub repository https://github.com/lxasqjc/learn-downsample, which provides the source code for the described methodology.
Open Datasets Yes We evaluate the performance of our deformed downsampling approach on three datasets from different domains as summarised in Table 1, against three baselines. Datasets are Cityscape (Cordts et al., 2016) and Deep Globe (Demir et al., 2018).
Dataset Splits Yes The Cityscapes (Cordts et al., 2016) dataset contains 5000 high-resolution (2048 x 1024 pixels) urban scenes images collected across 27 European Cities. The 5000 images from the Cityscapes are divided into 2975/500/1525 images for training, validation and testing. The Deep Globe (Demir et al., 2018) dataset has 803 high-resolution (2448 x 2448 pixels) images of aerial scenes...We randomly split the dataset into the train, validate and test with 455, 207, and 142 images respectively. The PCa-Histo: dataset contains 266 ultra-high resolution whole slide images...We random split the dataset into 200 training, 27 validation and 39 test images.
Hardware Specification Yes All networks are trained on 2 GPUs from an internal cluster (machine models: gtx1080ti, titanxp, titanx, rtx2080ti, p100, v100, rtx6000), with sync BN.
Software Dependencies No The paper mentions software like Adam and specific loss functions, but does not provide specific version numbers for any software dependencies (e.g., PyTorch version, Python version, CUDA version).
Experiment Setup Yes We optimize parameters using Adam (Kingma & Ba, 2014) with an initial learning rate of 1e-3 and β =0.9, and train for 200 epochs on Deep Globe, 250 epochs on PCa-Histo dataset, and 125 epochs on Cityscapes dataset. We use a batch size of 4 for the Cityscapes dataset (Cordts et al., 2016) and the Deep Globe dataset (Demir et al., 2018) and a batch size of 2 for the PCa-Histo dataset.