3 cocodetection mask implement #36

Johansmm · 2021-09-07T13:47:59Z

New features:

Include Segementation and append_segmentation attribute/function in Frame class
Segementation attribute in Frame: To work with segmentation datasets by binary features maps.
- Update Mask class Include get_view function to render segmentation
Use return_mask propierty of COCODetectionDataset to return a segmentation tensor.
Implement stuff_ann_file attribut of COCODetectionDataset to provide the stuff annotation json file (complement the things classes).
Integrate stuff classes and thing classes in COCODetectionDataset

…aset

thibo73800 · 2021-09-08T06:07:07Z

alodataset/coco_detection_dataset.py

@@ -16,7 +17,7 @@
 from torchvision.datasets.coco import CocoDetection

 from alodataset import BaseDataset
-from aloscene import BoundingBoxes2D, Frame, Labels
+from aloscene import BoundingBoxes2D, Frame, Labels, Panoptic


Wonder if Panoptic is the right class name. Panoptic refer to the task it self (stuff + things) while this tensor actually refer to the data. In imgaug, they refer to this type of class as : Segmentation Maps. Maybe Segmentation instead of Panoptic ? What do you think ?

Ok, you are right. I will change the name

alodataset/coco_detection_dataset.py

thibo73800 · 2021-09-08T06:16:39Z

alodataset/coco_detection_dataset.py

        frame.append_boxes2d(boxes)

+        if self.prepare.return_masks:
+            mask = Panoptic(target["masks"], names=("C", "H", "W"), labels=labels_2d)


If my understanding is correct the structure is the following:
Frame
- Boxes 2D
- label
- Panoptic
- label

There is therefore a constraints panoptic.shape[0] must be equal to boxes.shape[0] right ?
Is there any reason why you choose to not use one mask per boxes as you previously mentioned ?

Let's talk about it during the day

Yes. len(labels) == len(Boxes 2D) and there is a constraints with len(labels) == len(Panoptic). So len(Panoptic) == len(Boxes 2D).

The reason is used Segementation in applications that not required boxes.

thibo73800 · 2021-09-08T06:18:07Z

alodataset/coco_detection_dataset.py

-        level=logging.INFO,
-        format="[%(asctime)s][%(levelname)s] %(message)s",
-        datefmt="%d-%m-%y %H:%M:%S",
+        level=logging.INFO, format="[%(asctime)s][%(levelname)s] %(message)s", datefmt="%d-%m-%y %H:%M:%S",


Its getting weird to use logger in this main() method while we don't use it in other main methods in the datasets folder. I would suggest to remove it from here, removing the print, and use the sample as for the others datasets.

thibo73800 · 2021-09-08T06:20:11Z

aloscene/panoptic.py

+    """
+
+    @staticmethod
+    def __new__(cls, x: Tensor, labels: Union[dict, Labels] = None, names=("N", "H", "W"), *args, **kwargs):


We must probably describe what type of information labels is supposed to store. Before to merge,I would also suggest to add the docstrings in this class.

Panoptic was replaced by update of mask class

Johansmm · 2021-09-08T15:09:33Z

@thibo73800 All changes were made in the update of the mask class. Ready for revision.

thibo73800 · 2021-09-09T05:58:26Z

aloscene/frame.py

@@ -142,6 +138,19 @@ def append_disparity(self, disparity, name=None):
        """
        self._append_label("disparity", disparity, name)

+    def append_segmentation(self, segmentation: Mask, name: str = None):
+        """Attach a mask to the frame.


Can we replace mask by segmentation in the docstring ? Also we can add some information to explained why we rely on the Mask augmented tensor to add segmentation.

Ok. I updated the info as follows:

Parameters ---------- segmentation: aloscene.Mask Mask with size (N,H,W), where N is the features maps, each one for one object. Each feature map must be a binary mask. For that, is a type of aloscene.Mask

thibo73800 · 2021-09-09T06:00:30Z

aloscene/mask.py

        tensor = super().__new__(cls, x, *args, **kwargs)
+        tensor.add_label("labels", labels, align_dim=["N"], mergeable=True)
+        tensor.add_property("mask_size", x.shape)


This line does not seem great to me. If the frame get resized, the mask will be resized and this property might not be updated properly. Also it is still possible to get the shape of the mask at any time using self.shape

Done. Property removed.

thibo73800 · 2021-09-09T06:05:35Z

aloscene/mask.py

        tensor = super().__new__(cls, x, *args, **kwargs)
+        tensor.add_label("labels", labels, align_dim=["N"], mergeable=True)
+        tensor.add_property("mask_size", x.shape)
+        tensor.add_property("is_mask", not isinstance(labels, Labels))


Does is_mask make sense for a class with the name Mask ?

The following is weird

mask = Mask() mask.is_mask # False

OK. I removed this property also

Johansmm · 2021-09-09T07:36:46Z

@thibo73800 Ready for review again.

thibo73800

Sounds good

Johan MEJIA added 3 commits September 7, 2021 15:38

New features: Panotic as aloscene, segmentation and stuff in COCO dat…

959dfdf

…aset

Merge branch 'master' into 3-cocodetection-mask-implement

8e53575

Update cocodetection demo

f3d952f

Johansmm requested a review from thibo73800 September 7, 2021 13:48

Johansmm self-assigned this Sep 7, 2021

Johansmm added aloscene aloscene enhancement New feature or request labels Sep 7, 2021

thibo73800 requested changes Sep 8, 2021

View reviewed changes

Johan MEJIA added 3 commits September 8, 2021 16:54

Add segmentation attributo to frame

c889c83

Merge branch 'master' into 3-cocodetection-mask-implement

c03303a

Update COCO demo

ab8c15b

Johansmm requested a review from thibo73800 September 8, 2021 15:08

thibo73800 requested changes Sep 9, 2021

View reviewed changes

remove insignificant properties

2caad5d

Johansmm requested a review from thibo73800 September 9, 2021 07:36

thibo73800 approved these changes Sep 9, 2021

View reviewed changes

thibo73800 merged commit 04e8881 into master Sep 9, 2021

Johansmm deleted the 3-cocodetection-mask-implement branch October 20, 2021 08:01

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

3 cocodetection mask implement #36

3 cocodetection mask implement #36

Johansmm commented Sep 7, 2021 •

edited

Loading

thibo73800 Sep 8, 2021

Johansmm Sep 8, 2021

thibo73800 Sep 8, 2021

Johansmm Sep 8, 2021

thibo73800 Sep 8, 2021

Johansmm Sep 8, 2021

thibo73800 Sep 8, 2021

Johansmm Sep 8, 2021

Johansmm commented Sep 8, 2021

thibo73800 Sep 9, 2021

Johansmm Sep 9, 2021

thibo73800 Sep 9, 2021

Johansmm Sep 9, 2021

thibo73800 Sep 9, 2021

Johansmm Sep 9, 2021

Johansmm commented Sep 9, 2021

thibo73800 left a comment

3 cocodetection mask implement #36

3 cocodetection mask implement #36

Conversation

Johansmm commented Sep 7, 2021 • 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

Johansmm commented Sep 8, 2021

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

Johansmm commented Sep 9, 2021

thibo73800 left a comment

Choose a reason for hiding this comment

Johansmm commented Sep 7, 2021 •

edited

Loading