$ pip install -r requirements.txt
$ python kmeans_anchors_ratios.py --help
output:
usage: python kmeans_anchors_ratios.py \
--instances path/to/your_coco_instances.json \
--anchors-sizes 32 64 128 256 512 \
--input-size 512 \
--normalizes-bboxes True \
--num-runs 3 \
--num-anchors-ratios 3 \
--max-iter 300 \
--min-size 0 \
--iou-threshold 0.5 \
--decimals 1 \
--default-anchors-ratios '[(0.7, 1.4), (1.0, 1.0), (1.4, 0.7)]'
optional arguments:
-h, --help show this help message and exit
--instances N Path to the json instances file in COCO format.
--anchors-sizes N [N ...]
List of anchors sizes (e.g. [32, 64, 128, 256, 512]).
--input-size N Size according to which each image is resized before
being processed by the model.
--normalizes-bboxes N
Normalizes 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-anchors-ratios N
The number of anchors 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. min_size=32 implies that that all the
bounding boxes with an area less than or equal to 1024
(32 * 32) will be filtered. Default: 0.
--iou-threshold N Threshold above which anchors are assigned to ground-
truth object boxes. Default: 0.5.
--decimals N Number of decimals to use when rounding anchors
ratios. Default: 1.
--default-anchors-ratios N
List of anchors ratios to be compared with those found
by K-Means. It must be passed as a string, e.g.
'[(0.7, 1.4), (1.0, 1.0), (1.4, 0.7)]'. Default:
[(0.7, 1.4), (1.0, 1.0), (1.4, 0.7)].
$ wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip
$ unzip annotations_trainval2017.zip
$ python kmeans_anchors_ratios.py \
--instances ./annotations/instances_train2017.json \
--anchors-sizes 32 64 128 256 512 \
--input-size 512 \
--normalizes-bboxes True \
--num-runs 3 \
--num-anchors-ratios 3 \
--max-iter 300 \
--min-size 0 \
--iou-threshold 0.5 \
--decimals 1 \
--default-anchors-ratios '[(0.7, 1.4), (1.0, 1.0), (1.4, 0.7)]'
output:
[06/13 12:57:38] Reading ./annotations/instances_train2017.json
[06/13 12:57:54] Starting the calculation of the optimal anchors ratios
[06/13 12:57:54] Extracting and preprocessing bounding boxes
[06/13 12:57:57] Discarding 2 bounding boxes with size lower or equal to 0
[06/13 12:57:57] K-Means (3 runs): 100%|██████████████████| 3/3 [00:33<00:00, 11.06s/it]
Runs avg. IoU: 80.48% ± 0.00% (mean ± std. dev. of 3 runs, 0 skipped)
Avg. IoU between bboxes and their most similar anchors after norm. them to make their area equal (only ratios matter):
80.48%
[06/13 12:58:33] Default anchors ratios: [(0.7, 1.4), (1.0, 1.0), (1.4, 0.7)]
Avg. IoU between bboxes and their most similar default anchors, no norm. (both ratios and sizes matter): 55.16%
Num. bboxes without similar default anchors (IoU < 0.5): 253049/860001 (29.42%)
[06/13 12:58:37] K-Means anchors ratios: [(0.6, 1.5), (1.0, 1.0), (1.4, 0.7)]
Avg. IoU between bboxes and their most similar K-Means anchors, no norm. (both ratios and sizes matter): 55.72%
Num. bboxes without similar K-Means anchors (IoU < 0.5): 240788/860001 (28.00%)
[06/13 12:58:37] K-Means anchors have an IoU < 50% with bboxes in 1.43% less cases than the default anchors, you should consider to use them
For more infos see the tutorial.
- [2020-10-28] fixed error about EfficientDet0 anchors inside the tutorial (thanks @RomRoc) and added code for all EfficientDets
- [2020-06-13] added a comparision with the default anchors ratios, as suggested by @zylo117
- [2020-05-23] added avg. IoU between bounding boxes and anchors
- [2020-05-23] added a function to get annotations whose bounding boxes don't have similar anchors
- [2020-05-23] added a function to generate anchors given ratios and sizes
- [2020-05-17] created this repository
The code of this repo is mainly an adaptation of https://github.com/zhouyuangan/K-Means-Anchors.
Compared to the original, the K-Means implementation of this repo is more than 50x faster and some additional features are provided.