$ pip install -r requirements.txt
$ python kmeans_anchor_ratios.py --helpoutput:
usage: python kmeans_anchor_ratios.py \
--annotations-path path/to/your_coco_annotations.json \
--input-size 512 \
--scale-bboxes True \
--num-runs 10 \
--num-anchor-ratios 3 \
--max-iter 300 \
--min-size 0
optional arguments:
-h, --help show this help message and exit
--annotations-path N Path to the json annotation file in COCO format.
--input-size N Size to which each image is scaled before being
processed by the model.
--scale-bboxes N scale the bounding boxes, before giving them input to
K-Means, so that they have all an area of 1.
Default: True.
--num-runs N How many times to run K-Means. After the end of all
runs the best result is returned. Default: 1.
--num-anchor-ratios N
The number of anchor_ratios to generate. Default: 3.
--max-iter N Maximum number of iterations of the K-Means algorithm
for a single run. Default: 300.
--min-size N Size to which all bounding box sizes must be stricly
greater to be considered by K-Means. Filtering is
applied after rescaling the bounding boxes to the same
extent that the images are scaled to adapt them to the
input size. Default: 0.
$ wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip
$ unzip annotations_trainval2017.zip
$ python kmeans_anchor_ratios.py \
--annotations-path ./annotations/instances_train2017.json \
--input-size 512 \
--scale-bboxes True \
--num-runs 10 \
--num-anchor-ratios 3 \
--max-iter 300 \
--min-size 0output:
2020-05-17 17:30:09 Starting the calculation of the optimal anchor ratios
2020-05-17 17:30:09 Reading annotations from ./annotations/instances_train2017.json
2020-05-17 17:30:25 Extracting and preprocessing bounding boxes
2020-05-17 17:30:36 Discarding 2 bounding boxes with size lower or equal to 0
2020-05-17 17:30:37 K-Means (10 runs): 100%|████████████| 10/10 [01:10<00:00, 7.05s/it]
2020-05-17 17:31:47 Runs avg. IoU: 80.48% ± 0.00% (mean ± std. dev. of 10 runs, 0 skipped)
2020-05-17 17:31:47 Optimal anchor ratios (avg. IoU: 80.48%): [(0.6, 1.5), (1.0, 1.0), (1.4, 0.7)]
The code of this repo is mainly an adaptation of https://github.com/zhouyuangan/K-Means-Anchors.
Compared to the original, this code is faster and allows some additional settings.