Extended Mask R-CNN for RGB-D object instance segmentation based on Matterport's implementation. Segmentation accuracy improved by up to 31% using the additional depth input channel.
would it be possible to have a .h5 file uploaded with the weights obtained after training on the NYU dataset? im trying to train an RGB-D model on a small custom dataset and would rather not train a pretrained model on the NYU dataset from scratch, if possible
Thanks to share this code!
I'd like to follow training process.
In my understanding, this code(paper) uses image-net pretrained weight which has 3 channel input structure.
So I planed my training process as follows, roughly pretrained weight manipulate -> scenenet train ->custom data finetune.
Specifically
Load mask-rcnn image net pretrained weight
use module instance_segmentation/sceneNet/train.py with init_with == "rgb_to_rgbd"
then it results 4-channel input weight.
Train with init_with == "sceneNet"
Train with custom data
So my issue is, Is it right that init_with == "rgb_to_rgbd" part is the weight structure transforming part?
And, nowadays, Imagenet weight which attained model.load_weights(model.get_imagenet_weights(), by_name=True) may not works.
Hello, I'd like to use your code for instance segmentation at NYU Depth v2 dataset. But when I execute generate_np_data.py, the folder named 'data' made and there's no 'names' property at data folder. So I couldn't import this line, "from data.names import names". Could you tell me how to do this? Thank you.
My Error Message is this,
ModuleNotFoundError Traceback (most recent call last)
/tmp/ipykernel_3488/2422544616.py in
1 from instance_segmentation.objects_config import ObjectsConfig
2 from instance_segmentation.objects_dataset import ObjectsDataset, normalize
----> 3 from data.names import names
Hello. I'd like to train with scenenet dataset.
With your kind code, I followed to create training.pkl and validation.pkl.
But I can't find class_names.pkl. Let me know, please.
Thanks!