How to train MEGA on my own dataset？

[ZJU-lishuang]

No description provided.

[Scalsol]

You should organize your dataset like ImageNet VID and generate imageset files just like those in datasets/ILSVRC2015/ImageSets. For example, in VID_train_15frames.txt, every line contains 4 strings, each representing video folder, no meaning(but I still keep it), frame number, video length. You could check the dataset for further details.

[ZJU-lishuang]

where does it need to be modified if the number of categories has changed?

NUM_CLASSES in configs/BASE_RCNN_{}gpu.yaml
classes and classes_map in mega_core/data/datasets/vid.py

Does any other place need?

[Scalsol]

where does it need to be modified if the number of categories has changed?

NUM_CLASSES in configs/BASE_RCNN_{}gpu.yaml
classes and classes_map in mega_core/data/datasets/vid.py

Does any other place need?

I think it's enough.

[ZJU-lishuang]

How to test MEGA if current frame could only access information from previous frames？
Where does it need to be modified?

[Scalsol]

For local frame, change MEGA.MAX_OFFSET to zero, MEGA.KEY_FRAME_LOCATION to -MEGA.MIN_OFFSET and MEGA.ALL_FRAME_INTERVAL to -MEGA.MIN_OFFSET + 1. For global frame there is no elegant way to modify the current code to support that. One way to acheive this is to keep a queue to store all previous frame, and randomly select several of them when testing. Hope this helps.

[ZJU-lishuang]

Thanks for your help.I will have a try.

[HzZHoO]

No description provided.

Sorry to bother you. @ZJU-lishuang
May I ask you about the way that you annotate your own dataset? Does your dataset has annotations for every frame of the videos just like the VID dataset?

[ZJU-lishuang]

yes，the dataset is just like VID dataset.

[Dawn-LX]

where does it need to be modified if the number of categories has changed?

NUM_CLASSES in configs/BASE_RCNN_{}gpu.yaml
classes and classes_map in mega_core/data/datasets/vid.py

Does any other place need?

Thank you very much!, I only modified the classes and classes_map in mega_core/data/datasets/vid.py; and I encountered a RuntimeError:
RuntimeError: copy_if failed to synchronize: device-side assert triggered
CUDA error 59: Device-side assert triggered

Then I modified the NUM_CLASSES in configs/BASE_RCNN_{}gpu.yaml to the corresponding number of classes and the problem fixed

[Jezzzzz]

yes，the dataset is just like VID dataset.

Hi，how do you annotate your dataset like the VID dataset?

[ZJU-lishuang]

labelimg is enough