Training on custom dataset (Grayscale images)

[junxiant]

Hello everyone,

To train on custom datasets, i would like to check if what i am changing is correct. I would like some feedback or help please.

I have created my own custom dataset which is in grayscale (512x512), and i have converted it into the coco data format.

I'm using Mask-RCNN to do object detection and instance segmentation. Visualization using these coco json annotations seems correct, the segmentation and the bounding boxes matches the binary masks i have.

Since i am using "mask_rcnn_X_101_32x8d_FPN_3x.yaml" config, i have made the following changes:

1. PIXEL_MEAN: [0.485,0.456,0.406] (Might be wrong for grayscale)
2. PIXEL_STD: [0.229,0.224,0.225] (Again might be wrong for grayscale)
3. MODEL.ROI_HEADS.NUM_CLASSES: 5 (For 5 classes)
4. MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE: 64 (Reduced from original value of 512)
5. INPUT.FORMAT: 'L' (Grayscale)

Other settings such as:

1. SOLVER.IMS_PER_BATCH: 8
2. TEST.EVAL_PERIOD: 500
3. SOLVER.MAX_ITER: 3000
4. SOLVER.BASE_LR: 5e-3

When i trained it, the results were extremely bad, maybe i didnt train it long enough? Or maybe the pixel mean and std are wrong?
I would appreciate some feedback on these please.

"Bad results"

[junxiant]

Update:

I have commented out the PIXEL_MEAN and PIXEL_STD, and i think it definitely needs more training?