re: more precise segmentation

[dmarx]

just thought you might find these interesting if you hadn't seen them already:

• https://github.com/facebookresearch/Detic
• https://github.com/ashkamath/mdetr

[alexerupakat]

But what i understood is that doing a precise instance segmentation wouldn't give you a good result. I used mask generated out of instance segmentation and the result was pretty bad. i think it is because it needs a mask that is larger than the object. correct me if i am wrong.

[dmarx]

The way I've dealt with is by convolving the mask with a wide gaussian kernel, i.e. blur the edges of the mask. If you want a wider mask that's discrete valued, you could use this same procedure and just follow it up with a thresholding operation.

[alexerupakat]

The way I've dealt with is by convolving the mask with a wide gaussian kernel, i.e. blur the edges of the mask. If you want a wider mask that's discrete valued, you could use this same procedure and just follow it up with a thresholding operation.

thanks for the reply i will try that