ndsampler 0.5.13

Fast sampling from large images

The main webpage for this project is: https://gitlab.kitware.com/computer-vision/ndsampler

Fast random access to small regions in large images.

Random access is amortized by converting images into an efficient backend format (current backends include cloud-optimized geotiffs (cog) or numpy array files (npy)). If images are already in COG format, then no conversion is needed.

The ndsampler module was built with detection, segmentation, and classification tasks in mind, but it is not limited to these use cases.

The basic idea is to ensure your data is in MS-coco format, and then the CocoSampler class will let you sample positive and negative regions.

For classification tasks the MS-COCO data could just be that every image has an annotation that takes up the entire image.

Features

• CocoDataset for managing and manipulating annotated image datasets

• Amortized O(1) sampling of N-dimension space-time data (wrt to constant window size) (e.g. images and video).

• Hierarchical or mutually exclusive category management.

• Random negative window sampling.

• Coverage-based positive sampling.

• Dynamic toydata generator.

Also installs the kwcoco package and CLI tool.

Usage

The main pattern of usage is:

1. Use kwcoco to load a json-based COCO dataset (or create a kwcoco.CocoDataset programatically).

2. Pass that dataset to an ndsampler.CocoSampler object, and that effectively wraps the json structure that holds your images and annotations and it allows you to sample patches from those images efficiently.

3. You can either manually specify image + region or you can specify an annotation id, in which case it loads the region corresponding to the annotation.

Example

This example shows how you can efficiently load subregions from images.

>>> # Imagine you have some images
>>> import kwimage
>>> image_paths = [
>>>     kwimage.grab_test_image_fpath('astro'),
>>>     kwimage.grab_test_image_fpath('carl'),
>>>     kwimage.grab_test_image_fpath('airport'),
>>> ]  # xdoc: +IGNORE_WANT
['~/.cache/kwimage/demodata/KXhKM72.png',
 '~/.cache/kwimage/demodata/flTHWFD.png',
 '~/.cache/kwimage/demodata/Airport.jpg']
>>> # And you want to randomly load subregions of them in O(1) time
>>> import ndsampler
>>> # First make a COCO dataset that refers to your images (and possibly annotations)
>>> dataset = {
>>>     'images': [{'id': i, 'file_name': fpath} for i, fpath in enumerate(image_paths)],
>>>     'annotations': [],
>>>     'categories': [],
>>> }
>>> coco_dset = ndsampler.CocoDataset(dataset)
>>> print(coco_dset)
<CocoDataset(tag=None, n_anns=0, n_imgs=3, n_cats=0)>
>>> # Now pass the dataset to a sampler and tell it where it can store temporary files
>>> workdir = ub.ensure_app_cache_dir('ndsampler/demo')
>>> sampler = ndsampler.CocoSampler(coco_dset, workdir=workdir)
>>> # Now you can load arbirary samples by specifing a target dictionary
>>> # with an image_id (gid) center location (cx, cy) and width, height.
>>> target = {'gid': 0, 'cx': 200, 'cy': 200, 'width': 100, 'height': 100}
>>> sample = sampler.load_sample(target)
>>> # The sample contains the image data, any visible annotations, a reference
>>> # to the original target, and params of the transform used to sample this
>>> # patch
>>> print(sorted(sample.keys()))
['annots', 'im', 'params', 'tr']
>>> im = sample['im']
>>> print(im.shape)
(100, 100, 3)
>>> # The load sample function is at the core of what ndsampler does
>>> # There are other helper functions like load_positive / load_negative
>>> # which deal with annotations. See those for more details.
>>> # For random negative sampling see coco_regions.

A Note On COGs

COGs (cloud optimized geotiffs) are the backbone efficient sampling in the ndsampler library.

To preform deep learning efficiently you need to be able to effectively randomly sample cropped regions from images, so when ndsampler.Sampler (more acurately the FramesSampler belonging to the base Sampler object) is in “cog” mode, it caches all images larger than 512x512 in cog format.

I’ve noticed a significant speedups even for “small” 1024x1024 images. I haven’t made effective use of the overviews feature yet, but in the future I plan to, as I want to allow ndsampler to sample in scale as well as in space.

Its possible to obtain this speedup with the “npy” backend, which supports true random sampling, but this is an uncompressed format, which can require a large amount of disk space. Using the “None” backend, means that loading a small windowed region requires loading the entire image first (which can be ok for some applications).

Using COGs requires that GDAL is installed. Installing GDAL is a pain though.

https://gist.github.com/cspanring/5680334

Using conda is relatively simple

conda install gdal

# Test that this works
python -c "from osgeo import gdal; print(gdal)"

Also possible to use system packages

# References:
# https://gis.stackexchange.com/questions/28966/python-gdal-package-missing-header-file-when-installing-via-pip
# https://gist.github.com/cspanring/5680334


# Install GDAL system libs
sudo apt install libgdal-dev

GDAL_VERSION=`gdal-config --version`
echo "GDAL_VERSION = $GDAL_VERSION"
pip install --global-option=build_ext --global-option="-I/usr/include/gdal" GDAL==$GDAL_VERSION


# Test that this works
python -c "from osgeo import gdal; print(gdal)"

TODO

• [ ] Currently only supports image-based detection tasks, but not much work is needed to extend to video. The code was originally based on sampling code for video, so ndimensions is builtin to most places in the code. However, there are currently no test cases that demonstrate that this library does work with video. So we should (a) port the video toydata code from irharn to test ndcases and (b) fix the code to work for both still images and video where things break.

• [ ] Currently we are good at loading many small objects in 2d images. However, we are bad at loading images with one single large object that needs to be downsampled (e.g. loading an entire 1024x1024 image and downsampling it to 224x224). We should find a way to mitigate this using pyramid overviews in the backend COG files.