suoxinkey / PlenOctrees_NeRF-SH

PlenOctree Extraction algorithm
MIT License
58 stars 3 forks source link

Problem with mask and background #1

Open vahidEtt opened 3 years ago

vahidEtt commented 3 years ago

Hi,

I tried to train with sport_1 dataset, with below config, but it seems "USE_MASK" option doesn't work and I still see the background. Any idea what could be the problem?

SOLVER: OPTIMIZER_NAME: "Adam" BASE_LR: 0.0005 WEIGHT_DECAY: 0.0000000 IMS_PER_BATCH: 2 START_ITERS: 3000 END_ITERS: 60000 LR_SCALE: 0.1 WARMUP_ITERS: 0

MAX_EPOCHS: 100 CHECKPOINT_PERIOD: 3000 LOG_PERIOD: 30 BUNCH: 1000 COARSE_STAGE: 0

INPUT: SIZE_TRAIN: [256,192] SIZE_TEST: [256,192]

DATASETS: TRAIN: "/home/Downloads/sport_1" NUM_FRAME: 1 SHIFT: 0.0 MAXRATION: 0.0 ROTATION: 0.0 USE_MASK: True

DATALOADER: NUM_WORKERS: 8

MODEL: COARSE_RAY_SAMPLING: 64 FINE_RAY_SAMPLING: 128 SAMPLE_METHOD: "NEAR_FAR" # "NEAR_FAR" "BBOX" BOARDER_WEIGHT: 1e10 SAME_SPACENET: True TKERNEL_INC_RAW: True USE_SH: True DEVICE_IDS: [0]

TEST: IMS_PER_BATCH: 1

OUTPUT_DIR: "/media/Data_DL_HD/projects/PlenOctrees_NeRF-SH/output/sport_1"

Screenshot-20210416220214-1149x880

suoxinkey commented 3 years ago

Because the mask provided by the datasets is not so perfect. And We only sample the rays inside the mask during training and there is no supervision in the background area. so the above situation will occur.