Results are bad/random

[KoenL-1747481]

I have tried the demo file of your fork of the matterport mask rcnn repo and the resulting segmentations are very random : I experience the same issue with this tutorial. I also tried "augmented startups'" fork of the matterport mask rcnn repo (https://github.com/augmentedstartups/Mask_RCNN#readme) but running that demo also gives very random segmentations. It seems like the COCO pre-trained weights are just not working or something. Any idea how to fix this? Results from this tutorial look like this:

I tried this on 2 pc's with installations from scratch and both have the same random results.

[akTwelve]

It’s probably just drawing them wrong. Sometimes X and Y or Width and Height get swapped and the segmentations don’t get drawn correctly. I haven’t used this repo in a long time so I’m not exactly sure what’s up though.

[m11005217]

I have tried the demo file of your fork of the matterport mask rcnn repo and the resulting segmentations are very random : I experience the same issue with this tutorial. I also tried "augmented startups'" fork of the matterport mask rcnn repo (https://github.com/augmentedstartups/Mask_RCNN#readme) but running that demo also gives very random segmentations. It seems like the COCO pre-trained weights are just not working or something. Any idea how to fix this? Results from this tutorial look like this:

I tried this on 2 pc's with installations from scratch and both have the same random results.

Did you solve this problem, I also have same problem.

[KoenL-1747481]

I have tried the demo file of your fork of the matterport mask rcnn repo and the resulting segmentations are very random : I experience the same issue with this tutorial. I also tried "augmented startups'" fork of the matterport mask rcnn repo (https://github.com/augmentedstartups/Mask_RCNN#readme) but running that demo also gives very random segmentations. It seems like the COCO pre-trained weights are just not working or something. Any idea how to fix this? Results from this tutorial look like this: I tried this on 2 pc's with installations from scratch and both have the same random results.

Did you solve this problem, I also have same problem.

Yes and no.. i didn't fix it in this repo, but i stumbled upon facebook's mask rcnn implementation called detectron2 (search for the github repo). Its still actively being updated and it works perfectly with just a few lines of code (https://youtu.be/Pb3opEFP94U)

[m11005217]

Grateful ! Thanks for your help. I will do my best

[ansacaron]

Exactly the same problem here... any clue?

[KoenL-1747481]

Exactly the same problem here... any clue?

I moved on with facebook's mask rcnn implementation as i mentioned in a previous comment. Otherwise i have no clue

[pooyasohrabipoor]

It’s probably just drawing them wrong. Sometimes X and Y or Width and Height get swapped and the segmentations don’t get drawn correctly. I haven’t used this repo in a long time so I’m not exactly sure what’s up though.

Hi I am trying to use your code and your dataset but not only dont I get a good inference result, but also the validation loss is way higher than your results ( around 2 after 8 epochs) and when I use the inference my results are so random and bad. I do not want to use some tools like mmdetection because later on I need to modify maskrcnn code for my dataset. Do you have any idea or suggestions? I really appreciate your assistance.