English | Русский

LABELME-CONVERTER

Converter LabelMe to MsCOCO, PascalVOC, Yolo formats. Also, possible: splitting dataset into training, test and validation samples; augmentation of converted dataset; output and recording of statistics on images; transformation of the whole dataset to a single resolution.

Installing the necessary packages
>> pip install -r requirements.txt

The list of classes in the order they appear in the statistics is written in the .txt file (by default in labels.txt). The name of the file name is specified in config.py.

Converting

Start

The file labelme_converter.py with arguments is used for converting:

--input - directory with dataset in LabelMe format
--output - (optional) directory for storing converted dataset (default: ./current_dataset)
--format - format for conversion (yolo, voc, coco)
--create-marked - (optional) indicated if you need pictures with the markup visualization
--poly - (optional) specified if the visualization should be with polygons (not rectangles)

Example

Converting dataset from ./meters01_labelme to Yolo format with creation of markup visualization
>> python labelme_converter.py --input meters01_labelme --output current_dataset --format yolo --create-marked

Input path:  labelme
Output path:  current_dataset
Reading labelme files:
100%|██████████████████████████████████████████| 24/24 [00:00<00:00, 98.69it/s]
Create marked images:
100%|██████████████████████████████████████████| 24/24 [00:00<00:00, 50.44it/s]
Converting labelme to Yolo:
100%|██████████████████████████████████████████| 24/24 [00:00<00:00, 8058.87it/s]

Split on train, test, val and trainval

Start

The file split_dataset.py with arguments is used for partitioning:

--input - directory with dataset in one of the partitioning formats: MsCOCO, PascalVOC, Yolo
--output - (optional) directory for storing a split dataset (default: ./splitted_dataset)
--train - percentage of the total number of pictures which should get into the training set
--test - the percentage of the total number of pictures, which should go into the test set
--val - the percentage of the total number of pictures that should go into the validation set
--seed - (optional) value of the initial number of the random number generator (default: 42)

Example

Partitioning dataset from ./current_dataset in the ratio 80/15/5 with seed equal to 101
>> python split_dataset.py --input current_dataset --train 80 --test 15 --val 5 --seed 101

Dataset on current_dataset is being split!
Output split dataset path: splitted_dataset
Dataset is splitted!

Dataset augmentation

Start

The file augment_dataset.py with arguments is used for augmentation:

--input - a directory with a dataset in one of the markup formats: MsCOCO, PascalVOC, Yolo (dataset can be split)
--output - (optional) directory for storing a partitioned dataset (default: ./augmented_dataset)
--full - (optional) if you augment all sets in a split dataset, just specify this argument
--train - (optional) augmenting a training set in a split dataset
--test - (optional) augmentation of a test set in a split dataset
--val - (optional) augmentation of validation set in a split dataset
--count - number of augmented copies of one image

Examples

Augmentation of the entire unsplit set stored in ./current_dataset with 3 copies of each image
>> python augment_dataset.py --input current_dataset --count 3 

Augmentation train_test_val set:
100%|██████████████████████████████████████████| 24/24 [00:06<00:00,  3.52it/s]
(meters)

Augmentation of training and test sets from a split dataset (trainval set changes)
>> python augment_dataset.py --input splitted_dataset --train --test --count 5

Augmentation train set:
100%|██████████████████████████████████████████| 19/19 [00:07<00:00,  2.50it/s]
Augmentation test set:
100%|██████████████████████████████████████████| 3/3 [00:01<00:00,  2.62it/s]

Creating statistics on datasets

Start

The file statistics_dataset.py with arguments is used to output and record statistics:

--input - a directory with a dataset in one of the markup formats: MsCOCO, PascalVOC, Yolo (dataset can be split or augmented)
--save - (optional) if you want to save statistics to a file (stat.txt)
--save_path - (optional) if you want to save the file to a specific directory (by default it is saved in the dataset directory)

Examples

Display statistics for a converted dataset with saving to the file
>> python statistics_dataset.py --input current_dataset --save

+------------------------------------------------------------------+
|               train_test_val | count of images: 24               |
+---------+-------------------+-----------+------------+-----------+
|  class  | number of objects | avg_width | avg_height |  avg_area |
+---------+-------------------+-----------+------------+-----------+
|  meter  |         24        |   410.25  |   417.42   | 182052.79 |
|  value  |         24        |   191.29  |   51.21    |  11516.71 |
|  seal2  |         29        |   82.45   |   72.07    |  7140.14  |
|  model  |         17        |   119.24  |   26.94    |  3488.65  |
|  serial |         23        |   115.52  |   18.70    |  2652.00  |
|   seal  |         23        |   164.30  |   154.57   |  24377.17 |
|   mag   |         18        |   34.50   |   38.56    |  1516.61  |
| breaker |         16        |   155.88  |   196.06   |  31059.19 |
+---------+-------------------+-----------+------------+-----------+

Display statistics for a split dataset without saving to the file
>> python statistics_dataset.py --input splitted_dataset

+------------------------------------------------------------------+
|                   train | count of images: 19                    |
+---------+-------------------+-----------+------------+-----------+
|  class  | number of objects | avg_width | avg_height |  avg_area |
+---------+-------------------+-----------+------------+-----------+
|  meter  |         19        |   395.58  |   411.05   | 171485.26 |
|  value  |         19        |   183.05  |   51.11    |  10799.26 |
|   seal  |         16        |   155.19  |   151.06   |  21259.50 |
|   mag   |         13        |   30.62   |   33.54    |  1071.46  |
|  seal2  |         24        |   78.75   |   70.04    |  6600.00  |
|  model  |         14        |   107.64  |   25.93    |  2955.64  |
|  serial |         18        |   110.44  |   17.33    |  2226.56  |
| breaker |         12        |   160.50  |   184.33   |  31833.42 |
+---------+-------------------+-----------+------------+-----------+

+------------------------------------------------------------------+
|                    test | count of images: 3                     |
+---------+-------------------+-----------+------------+-----------+
|  class  | number of objects | avg_width | avg_height |  avg_area |
+---------+-------------------+-----------+------------+-----------+
|  meter  |         3         |   420.33  |   366.33   | 163746.67 |
|  value  |         3         |   217.33  |   47.00    |  13863.67 |
|   seal  |         5         |   146.20  |   151.40   |  22713.60 |
|   mag   |         3         |   38.33   |   50.67    |  2490.00  |
|  model  |         2         |   158.00  |   26.00    |  4578.00  |
|  serial |         3         |   123.33  |   19.33    |  3868.33  |
| breaker |         4         |   142.00  |   231.25   |  28736.50 |
|  seal2  |         3         |   65.33   |   52.67    |  3784.67  |
+---------+-------------------+-----------+------------+-----------+

+-----------------------------------------------------------------+
|                     val | count of images: 2                    |
+--------+-------------------+-----------+------------+-----------+
| class  | number of objects | avg_width | avg_height |  avg_area |
+--------+-------------------+-----------+------------+-----------+
| meter  |         2         |   534.50  |   554.50   | 309903.50 |
| value  |         2         |   230.50  |   58.50    |  14812.00 |
|  seal  |         2         |   282.50  |   190.50   |  53477.50 |
|  mag   |         2         |   54.00   |   53.00    |  2950.00  |
| seal2  |         2         |   152.50  |   125.50   |  18655.00 |
| serial |         2         |   149.50  |   30.00    |  4656.50  |
| model  |         1         |   204.00  |   43.00    |  8772.00  |
+--------+-------------------+-----------+------------+-----------+

+------------------------------------------------------------------+
|                  trainval | count of images: 21                  |
+---------+-------------------+-----------+------------+-----------+
|  class  | number of objects | avg_width | avg_height |  avg_area |
+---------+-------------------+-----------+------------+-----------+
|  meter  |         21        |   408.81  |   424.71   | 184667.95 |
|  value  |         21        |   187.57  |   51.81    |  11181.43 |
|   seal  |         18        |   169.33  |   155.44   |  24839.28 |
|   mag   |         15        |   33.73   |   36.13    |  1321.93  |
|  seal2  |         26        |   84.42   |   74.31    |  7527.31  |
|  model  |         15        |   114.07  |   27.07    |  3343.40  |
|  serial |         20        |   114.35  |   18.60    |  2469.55  |
| breaker |         12        |   160.50  |   184.33   |  31833.42 |
+---------+-------------------+-----------+------------+-----------+

Outputs statistics for an augmented unsplit dataset with the number of copies equal to 5. The statistics are also saved in the path ./stat/stat.txt
>> python statistics_dataset.py --input augmented_dataset --save --save_path stat

+------------------------------------------------------------------+
|              train_test_val | count of images: 120               |
+---------+-------------------+-----------+------------+-----------+
|  class  | number of objects | avg_width | avg_height |  avg_area |
+---------+-------------------+-----------+------------+-----------+
|  meter  |        120        |   426.29  |   441.95   | 201052.84 |
|  value  |        120        |   197.42  |   60.44    |  14066.77 |
|  seal2  |        145        |   85.43   |   74.40    |  7367.63  |
|  model  |         85        |   123.21  |   32.36    |  4433.02  |
|  serial |        115        |   118.58  |   23.93    |  3487.95  |
|   seal  |        115        |   172.86  |   164.56   |  27600.59 |
|   mag   |         90        |   35.52   |   41.13    |  1628.20  |
| breaker |         80        |   163.86  |   193.43   |  32002.40 |
+---------+-------------------+-----------+------------+-----------+

Transformation of all dataset images to the same resolution (with transformation of markup coordinates accordingly).

Start

The file resize_dataset.py with arguments is used for the transformation:

--input - directory with dataset in one of the markup formats: MsCOCO, PascalVOC, Yolo (dataset can be split or augmented)
--output - (optional) directory for saving a new dataset (default: ./resized_dataset)
--new_w - width of images in pixels
--new_h - image height in pixels

Example

Resize split dataset to 512x512 (in MsCOCO format three process bars are output for split dataset, in others - one)
Save to default directory: ./resized_dataset
>> python resize_dataset.py --input splitted_dataset --new_w 512 --new_h 512                        

Resize images in Train set:
100%|██████████████████████████████████████████| 19/19 [00:00<00:00, 27.08it/s]
Resize images in Test set:
100%|██████████████████████████████████████████| 3/3 [00:00<00:00, 25.54it/s]
Resize images in Val set:
100%|██████████████████████████████████████████| 2/2 [00:00<00:00, 28.34it/s]
Resize images in TrainVal set:
100%|██████████████████████████████████████████| 21/21 [00:00<00:00, 33.62it/s]

vladimirsinitsin / labelme_converter Goto Github PK