visual support for object detection dataset

[VaranRohila]

Added visualization support for object detection dataset.
Issue Link: Link

Procedure:

1. Find out all the keys with the Sequence feature connector (Bbox are wrapped inside them).
2. Find the keys with BboxFeature in Sequence keys.
3. Find the labels if there are any in the Sequence keys.
4. Plot image, bboxes, and labels on them.

We only take the first Sequence feature, first bbox feature and first label feature for now.

Notebook demonstration: Link

[Conchylicultor]

Thank you for this PR.

[VaranRohila]

@Conchylicultor I have made some major changes in the structure of the module as you had suggested. There is one main file visualization.py and two visualizer modules image_visualizer.py and object_visualizer.py

[Conchylicultor]

Thank you for the update. There seems to be a lot of duplicated code, making it harder to maintain. Could you re-factor in smaller functions reused across modules to avoid duplication ?

[VaranRohila]

@Conchylicultor I have made changes according to your suggestions. Created an ImageVisualizer base class and two child classes, SupervisedVisualizer and ObjectVisualizer.

Could you suggest some methods to reuse the code inside the build function? I was thinking of using python decorators but that would certainly increase the code complexity quite a bit.

[VaranRohila]

Update: I checked my code on wider_faces dataset as well. It is working fine, added that to the colab notebook for demonstration. I have also added a bbox_label flag to the parameters so that the label is not plotted for the bbox (for the datasets that have too many bbox and labels associated with them).

[VaranRohila]

@Conchylicultor Some structural changes in the code. The ImageVisulaizer now has a build function which acts as a binder for two functions, _infer_object_keys() and _plot_objects() which should be defined in the child classes.

In the _infer_object_keys() we retrieve keys of the objects we need to plot on a single image like for SupervisedVisualizer we retrieve label keys and ObjectVisualizer we retrieve the bboxes and the labels keys and we pass them to the _plot_objects() function in the child classes.

In the _plot_objects() method of the child classes, we plot the objects on a single example. We plot labels for SupervisedVisualizer; labels and objects for ObjectVisualizer.

[Conchylicultor]

Thank you for the update! This looks better

[VaranRohila]

@Conchylicultor I have also overridden build function in the child classes so that more arguments can be added with respect to different child classes like bbox_label flag in ObjectVisualizer.

[VaranRohila]

Also added DummyObjectDetection class in the test_utils.py for usage in the visualization_test.py file. Ran the test and it is working for both.

[Conchylicultor]

I ended up refactoring myself the visualization to make it more extensible: https://github.com/tensorflow/datasets/pull/1728/files
Could you refactor the code to use the new class ?

[VaranRohila]

Since I encountered some git errors, I had to force push. Everything is ok now/