robosat
robosat copied to clipboard
mapbox
Semantic segmentation on aerial and satellite imagery. Extracts features such as: buildings, parking lots, roads, water, clouds
For #167. Adds Feature Pyramid Attention (FPA) module :boom: :rocket: > Pyramid Attention Network for Semantic Segmentation https://arxiv.org/abs/1805.10180  _from https://arxiv.org/abs/1805.10180 Figure 2_  _from https://arxiv.org/abs/1805.10180 Figure 3_ Tasks -...
Similar to #187 With the introduction of the batching feature extraction (#148), the inputs to rs_subset is likely going to be multiple `.tiles` files. My current workaround to combine all...
At the moment the data loaders load up images from the dataset, do pre-processing (like normalization), and then convert the images into tensors. Then we copy the data from CPU...
For https://github.com/mapbox/robosat/issues/60. This changeset implements a Feature Pyramid Network (FPN) on top of a (potentially pre-trained) ResNet. - [Feature Pyramid Networks for Object Detection](https://arxiv.org/abs/1612.03144) - [A Unified Architecture for Instance...
For https://github.com/mapbox/robosat/issues/172 (see for context) - this implements the EfficientNetB0 model as an encoder for our encoder-decoder architecture. I'm currently training my EfficientNet model family (no h-swish, no squeeze-and-excitation) in...
This recently came out https://arxiv.org/abs/1905.11946 > EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks with incredible practical implications: we can either get a much better backbone than our current resnet50...
We should look into the feature pyramid attention (FPA) module for pixel-precise attention for segmentation features extracted from our resnet encoder. > Pyramid Attention Network for Semantic Segmentation https://arxiv.org/abs/1805.10180 ...
We should look into squeeze and excitation (SE) modules for adaptive feature map recalibration. The initially introduces cSE blocks are incredibly powerful for classification problems; the recently introduces sSE blocks...
Pytorch v1.1 (which we now use) comes with an official TensorBoard integration. We should remove our custom logging and chart plotting and move to TensorBoard. - https://pytorch.org/docs/stable/tensorboard.html - https://github.com/pytorch/pytorch/releases/tag/v1.1.0
A) MultiBands aka #56 - Switch from PIL to OpenCV for slippy map images (to allow multibands images handling) - SlippyMapTileConcatenation produce now an aggregate C,W,H NumPy tensor rather than...
