search

jiwoon-ahn / irn

Weakly Supervised Learning of Instance Segmentation with Inter-pixel Relations, CVPR 2019 (Oral)
MIT License
525 stars 100 forks source link

Asking about the Mask-Rcnn training strategy #11

Closed vicchu closed 5 years ago

vicchu commented 5 years ago

Hi, Jiwoon Ahn After transforming the pseudo label to the COCO-style annotations, I trained the Mask R-CNN with ResNet-50-FPN .

But the performance i got is slightly lower than the report ，mAP50 is 45.0.

image

I 'd like to ask you about the mask-rcnn training strategy, what kind data augmentation you adopt.

Thank you !

jiwoon-ahn commented 5 years ago

Hi @vicchu You may try with smaller image size and smaller training steps.

SOLVER: WEIGHT_DECAY: 0.0001 LR_POLICY: steps_with_decay BASE_LR: 0.01 GAMMA: 0.1 MAX_ITER: 18000 STEPS: [0, 12000, 16000] TRAIN: WEIGHTS: https://s3-us-west-2.amazonaws.com/detectron/ImageNetPretrained/MSRA/R-50.pkl DATASETS: ('voc_2012_train',) SCALES: (512,) MAX_SIZE: 800 BATCH_SIZE_PER_IM: 512 RPN_PRE_NMS_TOP_N: 2000 # Per FPN level IMS_PER_BATCH: 4 TEST: DATASETS: ('voc_2012_val',) SCALE: 512 MAX_SIZE: 800 NMS: 0.5 RPN_PRE_NMS_TOP_N: 1000 # Per FPN level RPN_POST_NMS_TOP_N: 1000 OUTPUT_DIR: .

vicchu commented 5 years ago

Hi @jiwoon-ahn After try the parameters you offered ,I still can not get the same result. But i have try larger image size,the result is closer to what the paper report.