Open MiaSanLei opened 3 years ago
l can almost obtain the video-map is about four points lower than the one reported in the paper, for ucf-24, video-mAP@0.5=44.96
Can you tell me what does your jhmdb experiment's cfg look like? My frame mAP is about 59% for jhmdb. ----------jhmdb.yaml---------- TRAIN:
DATASET: jhmdb21 # ava
, ucf24
or jhmdb21
BATCH_SIZE: 24
TOTAL_BATCH_SIZE: 128
LEARNING_RATE: 1e-4
EVALUATE: False
FINE_TUNE: False
BEGIN_EPOCH: 1
END_EPOCH: 10
SOLVER:
MOMENTUM: 0.9
WEIGHT_DECAY: 5e-4
STEPS: [3,4,5,6]
LR_DECAY_RATE: 0.5
ANCHORS: [0.95878, 3.10197, 1.67204, 4.0040, 1.75482, 5.64937, 3.09299, 5.80857, 4.91803, 6.25225]
NUM_ANCHORS: 5
OBJECT_SCALE: 5
NOOBJECT_SCALE: 1
CLASS_SCALE: 1
COORD_SCALE: 1
DATA:
NUM_FRAMES: 16
SAMPLING_RATE: 1
TRAIN_JITTER_SCALES: [256, 320]
TRAIN_CROP_SIZE: 224
TEST_CROP_SIZE: 224
MEAN: [0.4345, 0.4051, 0.3775]
STD: [0.2768, 0.2713, 0.2737]
MODEL:
NUM_CLASSES: 21
BACKBONE_3D: resnext101
BACKBONE_2D: darknet
WEIGHTS:
BACKBONE_3D: "weights/resnext-101-kinetics-hmdb51_split1.pth"
BACKBONE_2D: "weights/yolo.weights"
FREEZE_BACKBONE_3D: True
FREEZE_BACKBONE_2D: True
LISTDATA:
BASE_PTH: "/data1/su/datasets/JHMDB-YOWO"
TRAIN_FILE: "/data1/su/datasets/JHMDB-YOWO/trainlist.txt"
TEST_FILE: "/data1/su/datasets/JHMDB-YOWO/testlist.txt"
TEST_VIDEO_FILE: "/data1/su/datasets/JHMDB-YOWO/testlist_video.txt"
MAX_OBJS: 1
CLASS_NAMES: [
"brush_hair", "catch", "clap", "climb_stairs", "golf",
"jump", "kick_ball", "pick", "pour", "pullup", "push",
"run", "shoot_ball", "shoot_bow", "shoot_gun", "sit",
"stand", "swing_baseball", "throw", "walk", "wave"
]
BACKUP_DIR: "backup/jhmdb"
RNG_SEED: 1
NUM_GPUS: 4
VISBLE_GPUS: '"0, 1, 2, 3"'
GPUS_ID: [0, 1, 2, 3]
Inference with the best checkpoint file downloaded from the website. I can almost obtain the frame-mAP value reported in the paper, for ucf24, 80.37 vs 80.4, for jhmdb, 74.51 vs 74.4. But the video-map is about three points lower than the one reported in the paper, for jhmdb, video-mAP@0.5=82.57 vs 85.7. Could you tell me why, please?