Add VOC and SBD instance segmentation dataset #540

knorth55 · 2018-03-20T12:34:12Z

[EDIT] This PR is splitted into #540 and #541

add VOCInstanceSegmentationDataset
add SBDInstanceSegmenationDataset
add voc_instance_segmentation_label_names
add sbd_instance_segmentation_label_names
add assert_is_instance_segmentation_dataset
add test_voc_instance_segmentation_dataset.py
add test_sbd_instance_segmentation_dataset.py
add test_assert_is_instance_segmentation_dataset.py

knorth55 · 2018-03-20T12:45:48Z

[EDIT] This PR is splitted into #540 and #541

from chainercv.datasets import SBDInstanceSegmentationDataset
from chainercv.datasets import sbd_instance_segmentation_label_names
from chainercv.visualizations import vis_image
from chainercv.visualizations import vis_instance_segmentation
import matplotlib.pyplot as plt

dataset = SBDInstanceSegmentationDataset()
img, bbox, mask, label = dataset[0]
vis_instance_segmentation(
    img, bbox, mask, label, label_names=sbd_instance_segmentation_label_names,
    alpha=0.7)
plt.show()

yuyu2172 · 2018-03-20T13:08:05Z

chainercv/datasets/sbd/sbd_utils.py

+
+
+sbd_instance_segmentation_label_names = (
+    'background',


The data representation looks okay except for the range of label, which I want it to be [0, n_fg_class - 1]. n_fg_class is the number of classes excluding the background. This means that the label would never be "background". The same convention is adopted for detection.

I think that consistency really matters between detection and instance segmentation because two of them have very similar interface.

i add background as semantic segmentation because we normally have background label in instance segmentation.
The task is pixelwise segmentation and it means that label includes baxkground for all pixels.

now Mask-RCNN and FCIS have two stage structure with RPN, but this task predicts pixelwise probabilities, not detecting the object.

These methods are OK with no-background label names, so I updated to remove background label.
However, we need to think better way to merge semantic segmentation, instance segmentation and object detection.

yuyu2172 · 2018-03-20T13:12:25Z

chainercv/datasets/voc/voc_utils.py

@@ -62,6 +62,8 @@ def get_voc(year, split):
 voc_semantic_segmentation_label_names = (('background',) +
                                         voc_bbox_label_names)

+voc_instance_segmentation_label_names = voc_semantic_segmentation_label_names


yuyu2172 · 2018-03-20T13:13:37Z

chainercv/visualizations/vis_instance_segmentation.py

+from chainercv.visualizations.vis_semantic_segmentation import _default_cmap
+
+
+def vis_instance_segmentation(


I want this to be consistent with vis_bbox.
This means it should take img together with annotations.

i got comfused with vis_semantic_segmentation. why doesn't it have vis_image in vis_semantic _segmentation?

Good.

Here is a rationale behind the design.
In semantic segmentation, an annotation is displayed in a different plot from an image.
In the case of detection and instance segmentation, annotations are displayed on top of an image.
That is why taking image is mandatory for vis_instance_segmentation.

yuyu2172 · 2018-03-20T13:14:10Z

In [2]: chainercv.datasets.SBDInstanceSegmentationDataset()
---------------------------------------------------------------------------
FileNotFoundError                         Traceback (most recent call last)
<ipython-input-2-c6ccaee382d6> in <module>()
----> 1 chainercv.datasets.SBDInstanceSegmentationDataset()

~/projects/chainercv/chainercv/datasets/sbd/sbd_instance_segmentation_dataset.py in __init__(self, data_dir, split)
     32
     33         if data_dir == 'auto':
---> 34             data_dir = sbd_utils.get_sbd()
     35
     36         id_list_file = os.path.join(

~/projects/chainercv/chainercv/datasets/sbd/sbd_utils.py in get_sbd()
     17     if not os.path.exists(train_voc2012_file):
     18         six.moves.urllib.request.urlretrieve(
---> 19             train_voc2012_url, train_voc2012_file)
     20
     21     if os.path.exists(os.path.join(base_path, 'train.txt')):

~/miniconda2/envs/general3/lib/python3.6/urllib/request.py in urlretrieve(url, filename, reporthook, data)
    256         # Handle temporary file setup.
    257         if filename:
--> 258             tfp = open(filename, 'wb')
    259         else:
    260             tfp = tempfile.NamedTemporaryFile(delete=False)

FileNotFoundError: [Errno 2] No such file or directory: '/Users/yuyu/.chainer/dataset/pfnet/chainercv/sbd/benchmark_RELEASE/dataset/train_voc2012.txt'

Does downloader work?

EDIT:
You need to make the directory before calling urlretrieve if it does not exist.

yuyu2172 · 2018-03-20T13:20:35Z

chainercv/datasets/sbd/sbd_instance_segmentation_dataset.py

+            self.data_dir, 'cls', data_id + '.mat')
+        ins_file = os.path.join(
+            self.data_dir, 'inst', data_id + '.mat')
+        seg_mat = scipy.io.loadmat(seg_file)


Can you import scipy like this?
https://github.com/chainer/chainercv/blob/master/chainercv/visualizations/vis_image.py#L19

OK, i will add warnings too

yuyu2172 · 2018-03-20T13:21:57Z

setup.py

@@ -23,7 +23,8 @@
 setup_requires = ['numpy']
 install_requires = [
    'chainer>=3.2',
-    'Pillow'
+    'Pillow',
+    'scipy'


Could you not add this to the requirement?
I want to leave scipy as an optional dependency like matplotlib.

Please edit README.md and environment.yml instead.

knorth55 · 2018-03-20T16:35:09Z

You need to make the directory before calling urlretrieve if it does not exist.

thx, I updated.

knorth55 · 2018-03-20T18:24:31Z

@yuyu2172 I update to follow your advice and add test.

yuyu2172

Please add the following to docs (docs/source/reference)

assert_is_instance_segmentation_dataset
SBDInstanceSegmentationDataset
VOCInstanceSegmentationDataset
vis_instance_segmentation

yuyu2172 · 2018-03-21T01:40:42Z

chainercv/utils/testing/assertions/assert_is_instance_segmentation_dataset.py

+        'The shape of mask must be (N, H, W).'
+    assert label.shape == (N,), \
+        'The shape of label must be (N, ).'
+    assert label.min() >= -1 and label.max() < n_class, \


Will label.min() ever be ==-1?
If not, can you change the minimum allowed value to 0?

yuyu2172 · 2018-03-21T01:42:11Z

chainercv/utils/testing/assertions/assert_is_instance_segmentation_dataset.py

+
+    assert_is_image(img, color=True)
+    _, H, W = img.shape
+    N = bbox.shape[0]


Could you use R instead of N so that the notation is consistent with detection?

yuyu2172 · 2018-03-21T01:44:29Z

chainercv/datasets/sbd/sbd_utils.py

+    return base_path
+
+
+sbd_instance_segmentation_label_names = (


Since this is the same as voc_bbox_label_names, how about sbd_instance_segmentation_label_names = voc_instance_segmentation_label_names?

This way, we know that the two are the same.

yuyu2172 · 2018-03-21T01:45:45Z

chainercv/utils/testing/assertions/assert_is_instance_segmentation_dataset.py

+
+    assert isinstance(label, np.ndarray), \
+        'label must be a numpy.ndarray.'
+    assert bbox.dtype == np.float32, \


How about using assert_is_bbox?

yuyu2172 · 2018-03-21T01:46:22Z

chainercv/utils/testing/assertions/assert_is_instance_segmentation_dataset.py

+from chainercv.utils.testing.assertions.assert_is_image import assert_is_image
+
+
+def assert_is_instance_segmentation_dataset(dataset, n_class, n_example=None):


Can you rename this to n_fg_class?

yuyu2172 · 2018-03-21T01:46:35Z

chainercv/utils/testing/assertions/assert_is_instance_segmentation_dataset.py

+
+    Args:
+        dataset: A dataset to be checked.
+        n_class (int): The number of classes including background.


Change this to n_fg_class (int): The number of foreground classes.

yuyu2172 · 2018-03-21T01:50:38Z

chainercv/datasets/sbd/sbd_instance_segmentation_dataset.py

+    """Instance segmentation dataset for Semantic Boundaries Dataset `SBD`_.
+
+    The class name of the label :math:`l` is :math:`l` th element of
+    :obj:`chainercv.datasets.voc_semantic_segmentation_label_names`.


yuyu2172 · 2018-03-21T01:55:51Z

chainercv/datasets/voc/voc_instance_segmentation_dataset.py

+    """Instance segmentation dataset for PASCAL `VOC2012`_.
+
+    The class name of the label :math:`l` is :math:`l` th element of
+    :obj:`chainercv.datasets.voc_semantic_segmentation_label_names`.


yuyu2172 · 2018-03-21T01:56:27Z

chainercv/datasets/sbd/sbd_instance_segmentation_dataset.py

+
+        Returns:
+            tuple of color image, bounding boxes, masks and labels whose
+            shapes are (3, H, W), (N, 4), (N, H, W), (N, ) respectively.


Can you use R instead of N?
We have been using math notation for shapes.

:math:`R`

yuyu2172 · 2018-03-21T01:57:08Z

chainercv/datasets/sbd/sbd_instance_segmentation_dataset.py

+            H and W are height and width of the images, and N is the number
+            of objects in the image. The dtype of the color image and
+            the bounding boxes are :obj:`numpy.float32`, and that of the
+            masks and the labels are :obj:`numpy.int32`.


mask is not numpy.int32

yuyu2172 · 2018-03-21T01:57:56Z

chainercv/visualizations/vis_instance_segmentation.py

+        >>> plot.show()
+
+    Args:
+        img (~numpy.ndarray): An array of shape :math:`(3, height, width)`.


height, width --> H, W

yuyu2172 · 2018-03-21T01:58:05Z

chainercv/visualizations/vis_instance_segmentation.py

+            represents a bounding box of an object.
+            The bounding box is :math:`(y_min, x_min, y_max, x_max)`.
+        mask (~numpy.ndarray): A bool array of shape
+            :math`(N, height, width)`.


yuyu2172 · 2018-03-21T01:59:36Z

tests/visualizations_tests/test_vis_instance_segmentation.py

+
+    def setUp(self):
+        self.img = np.random.randint(0, 255, size=(3, 32, 48))
+        self.mask = np.random.randint(0, 1, size=(self.n_bbox, 32, 48))


This is always 0.
np.random.randint(0, 2, size= ..., dtype=np.bool) is better.

yuyu2172 · 2018-03-21T01:59:44Z

tests/visualizations_tests/test_vis_instance_segmentation.py

+
+    def setUp(self):
+        self.img = np.random.randint(0, 255, size=(3, 32, 48))
+        self.mask = np.random.randint(0, 1, size=(self.n_bbox, 32, 48))


yuyu2172 · 2018-03-21T02:00:34Z

Please add tests for assert_is_instance_segmentation_dataset.

yuyu2172 · 2018-03-21T02:00:53Z

I think it is better to put vis_instance_segmentation into a different PR.

knorth55 · 2018-03-21T04:09:32Z

@yuyu2172 Updated and split this PR into two (this and #541).

yuyu2172 · 2018-03-21T11:17:56Z

chainercv/utils/testing/assertions/assert_is_instance_segmentation_dataset.py

+    img, bbox, mask, label = example[:4]
+
+    assert_is_image(img, color=True)
+    assert_is_bbox(bbox)


Can you pass size argument?

yuyu2172 · 2018-03-21T11:19:00Z

chainercv/utils/testing/assertions/assert_is_instance_segmentation_dataset.py

+        'The type of bbox must be bool'
+    assert label.dtype == np.int32, \
+        'The type of label must be numpy.int32.'
+    assert (bbox[:, 0] < bbox[:, 2]).all(), \


These assertions for bbox is not necessary because assert_is_bbox is used.

yuyu2172 · 2018-03-21T11:19:16Z

chainercv/utils/testing/assertions/assert_is_instance_segmentation_dataset.py

+    R = bbox.shape[0]
+
+    assert isinstance(label, np.ndarray), \
+        'label must be a numpy.ndarray.'


assert similar test for mask

yuyu2172 · 2018-03-21T11:19:43Z

chainercv/utils/testing/assertions/assert_is_instance_segmentation_dataset.py

+
+    assert isinstance(label, np.ndarray), \
+        'label must be a numpy.ndarray.'
+    assert mask.dtype == bool, \


bool -> np.bool

Also, the message should be The type of mask must be np.bool..

yuyu2172 · 2018-03-21T11:20:51Z

...s/utils_tests/testing_tests/assertions_tests/test_assert_is_instance_segmentation_dataset.py

+    def get_example(self, i):
+        img = np.random.randint(0, 256, size=(3, 48, 64))
+        n_bbox = np.random.randint(10, 20)
+        mask = np.random.randint(0, 1, size=(n_bbox, 48, 64), dtype=bool)


randint(0, 2, ...)

yuyu2172 · 2018-03-21T11:38:28Z

chainercv/datasets/sbd/sbd_instance_segmentation_dataset.py

+            mask.append(mask_inst)
+            label.append(instance_class)
+        bbox = np.array(bbox).astype(np.float32)
+        mask = np.array(mask).astype(bool)


Please use np.bool instead of bool.
I prefer not to use Python's builit-in types for casting numpy arrays because it is inconsistent.
We do not do array.astype(float), but do array.astype(np.float64).

I think np.bool is same as bool.
In numpy doc, there is no np.bool dtype officially.
https://docs.scipy.org/doc/numpy-1.13.0/user/basics.types.html

In [1]: import numpy as np In [2]: np.bool Out[2]: bool In [3]: bool Out[3]: bool In [4]: bool == np.bool Out[4]: True

yuyu2172 · 2018-03-21T11:38:46Z

chainercv/datasets/voc/voc_instance_segmentation_dataset.py

+            mask.append(mask_inst)
+            label.append(instance_class)
+        bbox = np.array(bbox).astype(np.float32)
+        mask = np.array(mask).astype(bool)


yuyu2172 · 2018-03-21T11:44:32Z

chainercv/datasets/sbd/sbd_instance_segmentation_dataset.py

+            assert inst_id != -1
+            assert instance_class != -1
+
+            where = np.argwhere(mask_inst)


inst, ins_*, instances are used.
Please be consistent.

I would like to make a suggestion.

instances --> inst_ids

ins_img --> inst_img

ins_file --> inst_file

Also, it is weird that instance_class is an element of label. Same goes for mask_inst.
How about

instance_class --> lbl or just label.append(np.unique(seg_img[mask_inst])[0] - 1)

mask_inst --> m

I prefer inst_*, so I will update.

yuyu2172 · 2018-03-21T11:47:09Z

chainercv/datasets/sbd/sbd_instance_segmentation_dataset.py

+            assert instance_class != -1
+
+            where = np.argwhere(mask_inst)
+            (y1, x1), (y2, x2) = where.min(0), where.max(0) + 1


y1,x1,y2,x2 -> y_min, x_min, y_max, x_max
This is the notation we use.

yuyu2172 · 2018-03-21T11:47:29Z

chainercv/datasets/voc/voc_instance_segmentation_dataset.py

+            assert instance_class != -1
+
+            where = np.argwhere(mask_inst)
+            (y1, x1), (y2, x2) = where.min(0), where.max(0) + 1


yuyu2172 · 2018-03-21T13:19:26Z

chainercv/datasets/sbd/sbd_instance_segmentation_dataset.py

+        ins_img = ins_mat['GTinst']['Segmentation'][0][0].astype(np.int32)
+        ins_img[ins_img == 0] = -1
+        ins_img[ins_img == 255] = -1
+        return seg_img, ins_img


seg_img is a short notation for "segmentation image", but ins_img is also an image with segmentations.
How about label_img?

the value of ins_img is inst_id, not label.
label_img is not a good name, i think.
I prefer inst_img.

Sorry. I was talking about chaning seg_img.
My proposal is to use label_img instead of seg_img. inst_img looks good.

Something like this

seg_file --> label_file.

seg_anno --> label_anno.

seg_img --> label_img

OK, I will update

knorth55 · 2018-03-22T00:35:43Z

@yuyu2172
I searched a little bit about numpy boolean datatype.
I think numpy.bool_ is the correct dtype.
https://docs.scipy.org/doc/numpy-1.13.0/reference/generated/numpy.dtype.html

In [1]: import numpy as np

In [2]: a = np.array([True, False], dtype=bool)

In [3]: a.dtype.type
Out[3]: numpy.bool_

In [4]: a = np.array([True, False], dtype=np.bool)

In [5]: a.dtype.type
Out[5]: numpy.bool_

In [6]: a = np.array([True, False], dtype=np.bool_)

In [7]: a.dtype.type
Out[7]: numpy.bool_

In [8]: np.bool_ == bool
Out[8]: False

In [9]: np.bool_ == np.bool
Out[9]: False

In [10]: bool == np.bool
Out[10]: True

yuyu2172 · 2018-03-22T03:41:15Z

chainercv/utils/testing/assertions/assert_is_instance_segmentation_dataset.py

+    assert_is_image(img, color=True)
+    assert_is_bbox(bbox, size=(H, W))
+
+    assert isinstance(bbox, np.ndarray), \


This is redundant.

yuyu2172 · 2018-03-22T03:43:38Z

chainercv/datasets/sbd/sbd_instance_segmentation_dataset.py

+        ins_img = ins_mat['GTinst']['Segmentation'][0][0].astype(np.int32)
+        ins_img[ins_img == 0] = -1
+        ins_img[ins_img == 255] = -1
+        return seg_img, ins_img


Sorry. I was talking about chaning seg_img.
My proposal is to use label_img instead of seg_img. inst_img looks good.

Something like this

seg_file --> label_file.

seg_anno --> label_anno.

seg_img --> label_img

knorth55 · 2018-03-23T07:19:39Z

@yuyu2172 Updated to pass the test,

yuyu2172

You do not need to rebase commits
(IMO rebasing is not recommend because the change from the previously reviewed version is not clear)

yuyu2172 · 2018-03-24T12:56:22Z

chainercv/datasets/voc/voc_instance_segmentation_dataset.py

+        inst_img[inst_img == 255] = -1
+        return label_img, inst_img
+
+    def _prepare_data(self, label_img, inst_img):


If you want to reuse this method, how about putting this method to voc_utils?
I did not find inheritance relationship betweenVOC and SBD effective.

The name of the method should be something like
convert_to_instance_wise_anno

Maybe convert_to_instance_segmentation is better

Or image_wise_to_instance_wise

This name is nice because the name of the opposite operation is obvious.

If these operations are common, we can support them under utilities similar to bbox2loc and loc2bbox

For now, you do not need to make the functions fully supported in utilities. This is just a possibility.

yuyu2172 · 2018-03-24T12:57:02Z

chainercv/datasets/voc/voc_instance_segmentation_dataset.py

+        bbox, mask, label = self._prepare_data(label_img, inst_img)
+        return img, bbox, mask, label
+
+    def _load_label_ins(self, data_id):


This should be _load_label_inst

knorth55 · 2018-03-25T11:17:25Z

I rebase the commits because it is easy to understand from git log.
Update or fix or some other small commits are pretty annoying, and I dont expect this kind of endless review.

yuyu2172 · 2018-03-25T11:58:25Z

I rebase the commits because it is easy to understand from git log.

OK. I am just suggesting that rebasing is not necessary just in case you do not know. Please feel free to choose whatever you need to do. And yes, this PR is big.

yuyu2172 · 2018-03-26T02:27:22Z

Please resolve conflict with the master branch.

knorth55 · 2018-03-26T04:49:23Z

I resolved conflicts and update to follow the reviews.

knorth55 requested a review from yuyu2172 March 20, 2018 12:34

yuyu2172 mentioned this pull request Mar 20, 2018

Add Instance Segmentation #538

Open

21 tasks

knorth55 force-pushed the add-voc-instance-segmentation-dataset branch from 71d5e27 to 800f71e Compare March 20, 2018 12:41

knorth55 changed the title ~~Add voc instance segmentation dataset~~ Add VOC and SBD instance segmentation dataset and visualize function Mar 20, 2018

knorth55 added the feature label Mar 20, 2018

knorth55 self-assigned this Mar 20, 2018

yuyu2172 reviewed Mar 20, 2018

View reviewed changes

knorth55 force-pushed the add-voc-instance-segmentation-dataset branch from 2f22f84 to f2f3f9e Compare March 20, 2018 16:33

knorth55 changed the title ~~Add VOC and SBD instance segmentation dataset and visualize function~~ [WIP] Add VOC and SBD instance segmentation dataset and visualize function Mar 20, 2018

knorth55 changed the title ~~[WIP] Add VOC and SBD instance segmentation dataset and visualize function~~ Add VOC and SBD instance segmentation dataset and visualize function Mar 20, 2018

yuyu2172 reviewed Mar 21, 2018

View reviewed changes

knorth55 force-pushed the add-voc-instance-segmentation-dataset branch 2 times, most recently from 5c0bcb1 to 69c2402 Compare March 21, 2018 03:57

knorth55 changed the title ~~Add VOC and SBD instance segmentation dataset and visualize function~~ Add VOC and SBD instance segmentation dataset Mar 21, 2018

knorth55 mentioned this pull request Mar 21, 2018

Add vis instance segmentation func #541

Merged

2 tasks

knorth55 force-pushed the add-voc-instance-segmentation-dataset branch from 69c2402 to 8bf9cc2 Compare March 21, 2018 04:07

knorth55 force-pushed the add-voc-instance-segmentation-dataset branch 2 times, most recently from 2020492 to 484e239 Compare March 21, 2018 04:35

yuyu2172 reviewed Mar 21, 2018

View reviewed changes

knorth55 force-pushed the add-voc-instance-segmentation-dataset branch from 484e239 to bd11c7a Compare March 22, 2018 00:20

knorth55 force-pushed the add-voc-instance-segmentation-dataset branch 2 times, most recently from d180138 to a05cc2e Compare March 22, 2018 00:40

yuyu2172 reviewed Mar 22, 2018

View reviewed changes

knorth55 force-pushed the add-voc-instance-segmentation-dataset branch 3 times, most recently from 9529629 to 59c0756 Compare March 23, 2018 05:31

knorth55 force-pushed the add-voc-instance-segmentation-dataset branch from 59c0756 to 0c04ffe Compare March 23, 2018 14:37

yuyu2172 reviewed Mar 24, 2018

View reviewed changes

knorth55 added 10 commits March 26, 2018 13:48

add VOC instance segmentation dataset

c0ca2c0

add SBDInstanceSegmentationDataset

0ef54fe

add SciPy as additional dependencies

f3195b4

add assert_is_instance_segmentation_dataset.py

585aafb

add test_voc_instance_segmentation_dataset.py

4fbf2ef

add test_sbd_instance_segmentation_dataset.py

23ba7b6

add VOCInstanceSegmentationDataset doc

8aef84d

add SBDInstanceSegmentationDataset doc

4966475

add assert_is_instance_segmentation_dataset doc

5f4b5f0

add test_assert_is_instance_segmentation_dataset.py

08258bc

knorth55 force-pushed the add-voc-instance-segmentation-dataset branch from 0c04ffe to 08258bc Compare March 26, 2018 04:48

yuyu2172 approved these changes Mar 26, 2018

View reviewed changes

yuyu2172 merged commit ff7f705 into chainer:master Mar 26, 2018

knorth55 deleted the add-voc-instance-segmentation-dataset branch March 26, 2018 05:22

yuyu2172 added this to the v0.9 milestone Apr 17, 2018

		from chainercv.visualizations.vis_semantic_segmentation import _default_cmap


		def vis_instance_segmentation(

		from chainercv.utils.testing.assertions.assert_is_image import assert_is_image


		def assert_is_instance_segmentation_dataset(dataset, n_class, n_example=None):

Add VOC and SBD instance segmentation dataset #540

Add VOC and SBD instance segmentation dataset #540

Conversation

knorth55 commented Mar 20, 2018 • edited Loading

knorth55 commented Mar 20, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

yuyu2172 Mar 21, 2018 • edited Loading

Choose a reason for hiding this comment

yuyu2172 commented Mar 20, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

knorth55 commented Mar 20, 2018

knorth55 commented Mar 20, 2018

yuyu2172 left a comment • edited by knorth55 Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

yuyu2172 commented Mar 21, 2018

yuyu2172 commented Mar 21, 2018

knorth55 commented Mar 21, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

knorth55 Mar 21, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

knorth55 Mar 21, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

knorth55 commented Mar 22, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

knorth55 commented Mar 23, 2018

yuyu2172 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

yuyu2172 Mar 25, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

knorth55 commented Mar 25, 2018

yuyu2172 commented Mar 25, 2018

yuyu2172 commented Mar 26, 2018 • edited Loading

knorth55 commented Mar 26, 2018

knorth55 commented Mar 20, 2018 •

edited

Loading

knorth55 commented Mar 20, 2018 •

edited

Loading

yuyu2172 Mar 21, 2018 •

edited

Loading

yuyu2172 commented Mar 20, 2018 •

edited

Loading

yuyu2172 left a comment •

edited by knorth55

Loading

knorth55 commented Mar 21, 2018 •

edited

Loading

knorth55 Mar 21, 2018 •

edited

Loading

knorth55 Mar 21, 2018 •

edited

Loading

knorth55 commented Mar 22, 2018 •

edited

Loading

yuyu2172 Mar 25, 2018 •

edited

Loading

yuyu2172 commented Mar 26, 2018 •

edited

Loading