1

How can I improve the mask prediction by Mask RCNN model? Bounding box and class prediction seems to be okay in my case. Masks are kind of acceptable for small objects but not for big objects. The story is similar for other images as well. Here's my configurations:

RPN_ANCHOR_SCALES = (16, 32, 64, 128, 256)

TRAIN_ROIS_PER_IMAGE = 64

MAX_GT_INSTANCES = 50

POST_NMS_ROIS_INFERENCE = 500

POST_NMS_ROIS_TRAINING = 1000

USE_MINI_MASK True

MASK_SHAPE [28, 28]

MINI MASK_SHAPE [56, 56]

LEARNING_RATE = 0.001

LEARNING_MOMENTUM = 0.9

WEIGHT_DECAY = 0.0001

EPOCHS = 500

Any suggestions would be great!

results

DigitalAG
  • 41
  • 4

1 Answers1

1

Wanted to give an update to my post. I improved my accuracy by changing default mask_shape to [56, 56]. In order to be able to change the config, an extra conv2dtranspose layer should be added in model.train py

DigitalAG
  • 41
  • 4
  • But where specifically you add this layer? – Rishik Mani Feb 26 '21 at 14:04
  • @RishikMani, You need to add the layer in the line 1000 and 1001 in model.py 'x = KL.TimeDistributed(KL.Conv2DTranspose(256, (2, 2), strides=2, activation="relu"), name="mrcnn_mask_deconv1")(x) x = KL.TimeDistributed(KL.Conv2DTranspose(256, (2, 2), strides=2, activation="relu"), name="mrcnn_mask_deconv2")(x)' – DigitalAG Mar 01 '21 at 14:43