Semi-automatic tools enhancements (Stage 0) #3417
Conversation
bsekachev
changed the title
| max(bbox[1] - crop_padding, 0), | ||
| min(bbox[2] + crop_padding, image.width - 1), | ||
| min(bbox[3] + crop_padding, image.height - 1) | ||
| max(bbox[0][0] - crop_padding, 0), |
There was a problem hiding this comment.
Is it unified with some other input parameters? Usually a bounding box is 4 coordinates [x0, y0, x1, y1]. I'm OK with any solution if there is a common ground behind the decision.
There was a problem hiding this comment.
Also it will be necessary to redeploy the function with the change. Again, I'm fine with that if we have strong reasons to unify different interfaces.
There was a problem hiding this comment.
Done like for pos_points, neg_points.
We anyway need to re-deploy all the functions after adding min_neg_points
|
@bsekachev , why did you decide to keep bounding box till the object is finished? Can I resize it? Please look at the interface below. Probably we need something similar without the extra bbox during adding points. What do you think?
|
You can't resize it. I believe this bounding box does not disturb a user anyhow, maybe it even helps to remember what region was selected initially, but If you want it is removed, it can be done easily. |
I like the demo and interface from IOG github page. I will vote to remove the box. |
|
@bsekachev , are you going to unify interfaces to remove points with polygons in the future? |
Done. |
I am not sure they can be unified absolutely the same way because we must have negative points in the first case and we must keep dragging points in the second case. At least I think need to use an additional hotkey. |
| @@ -7,6 +7,7 @@ metadata: | |||
| spec: | |||
| framework: pytorch | |||
| min_pos_points: 1 | |||
| min_neg_points: 0 | |||
There was a problem hiding this comment.
It looks like the method can have negative points:
Clicks representations: We use the same clicks representation as DEXTR[46] by centering a 2D Gaussian around
each click, creating two separate heatmaps for foreground
and background clicks. The resulting heatmaps are concatenated with the RGB input image to form a 5-channel
input for the network. Similar to [46], the bounding box is
first relaxed by several pixels to include context, followed
by cropping to focus on the object-of-interest (Figure 2(d)).
There was a problem hiding this comment.
Although our IOG approach requires only three clicks
to perform segmentation, our framework naturally supports interactive adding of new foreground and background
clicks for further refinement if the user is not satisfied with
the current segmentation output. To achieve this, we append a lightweight branch before the PSP module to accept the two-channel Gaussian heatmaps encoding all the
foreground and background clicks
There was a problem hiding this comment.
Okay, it has in the implementation.
min_neg_poinst: 0 means that method supports negative points, but they are not obligatory.
|
Could you please update the documentation a bit?
If you need a sever with these changes, please contact @azhavoro |
|
@bsekachev We'll update the documentation |
|
Could you please update tests where we create polygons/polylines/points (need to test "Done" button) |
Motivation and context
Resolved #3361
Related #2936
Related #2515
How has this been tested?
Checklist
developbranchcvat-core, cvat-data and cvat-ui)
License
Feel free to contact the maintainers if that's a concern.