Getting less mAP on MMdetection

Hello, everyone. I spent a month looking for an appropriate single mmdet base model that would yield at least 30mAP, but all of these models yielded lower mAP values than detectron2 models. Except for training for extra epochs, I tried all the fundamental strategies like data augmentation and multiscale training , hyper parameter tuning (trained for 20 epochs). I even looked back at published previous challenge best solutions, and they took the similar pattern. If it’s okay , please provide any basic mmdet techniques like higher resolution, ensemble, bigger models, longer training periods that are working well for everyone but that I overlooked.
thanks.

I guess I know what top participants are using but I cannot reveal that at the moment since the competition is still ongoing. What I can tell you is:

  • I am highly confident that they are using MMDetection
  • You haven’t tried the previous challenge’s best solution. I know that because I won it, and as simple as Mask RCNN yielded slightly better performance than what you are getting now
  • Data augmentation, Multi-scale training do make a significant improvement
  • The hyper-paramters which are well known in image classification tasks play a role in instance segmentation but not too much, the research has its own hyper-parameters that you will find in test_config
  • Regarding ensemble. Yes, it is not working in this challenge because AP@50 is quite robust and naive ensembling will not be enough. You need some more post-processing steps to filter out false positives.
3 Likes

Hi, @saidinesh_pola,
There is a big secret here in data processing before training (as for me). And it is not so big difference between detectron2 and mmdetection.
I assumed that it is not big secret that
#1 team uses mmdet,
#2 team uses mmdet,
#3 team uses mmdet,
#4 team uses detectron2,
(you could observe it due to the errors in some of their submissions).
So, as for me mmdet is better due to large variety of different models, but it is not the main reason here.
For a long time I could not find the idea with data processing and tried to train data as is. And found that only 1 stage models could get good score, however all models with 2 and more stages are not good here (like mask-rcnn, HTC, detectors and so on.)
When I understood the difference between 1 and 2 stage models, I quickly find the data processing idea.
I think difference between your score and other teams from top is dataprocessing step, not the model, parameters or sophisticated augmentation.
I’d recommend you to try out the 1 stage model and see the results…

1 Like

If you need more info about difference between 1 and 2 stage models, you could read about RPN (region propose network) for example, how it works.
And also, for example about SOLO (1 stage model).
You could try solo in mmdet library as is and see if it helps you to boost your score.

2 Likes