Skip to content

Latest commit

 

History

History

instance_seg

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
lvis0.5/mask_rcnn/res50
lvis0.5/mask_rcnn/res50
 
 
lvis1.0/mask_rcnn/res50
lvis1.0/mask_rcnn/res50
 
 
README.md
README.md
 
 

README.md

Instructions for LVIS Dataset

Prepare Dataset

Please put the dataset and annotation into the cvpods project as following:

.
└── datasets
    ├── coco 
    │     ├── annotations
    │     ├── train2017
    │     └── val2017
    ├── lvis
    │     ├── lvis_v0.5_train.json
    │     ├── lvis_v0.5_val.json
    │     ├── num_shots_v0.5.npy
    │     ├── lvis_v1_train.json
    │     ├── lvis_v1_val.json
    │     └── num_shots_v1.0.json
    └── .....

Related files can be downloaded:

Training and Inference

  • Enter one project folder
  • Traning with:
pods_train --num-gpus 8
  • Inference with:
pods_test --num-gpus 8

Model Zoo

(* denotes the cosine classifier)

LVIS V0.5(Mask R-CNN)

We refactor the code of the internal version and re-train all experiments, the performance results have a little difference(higher) with the reported in the original paper.

ResNet-50

Name Cls Norm input size lr sched train time (s/iter) train mem (GB) box AP mask AP Trained Model
MaskRCNN-R50-FPN 640-800 90k 0.486 5.26 20.4 20.7 LINK
MaskRCNN-R50-FPN Cosine 640-800 90k 0.500 5.26 23.0 23.8 LINK
MaskRCNN-R50-FPN-RFS 640-800 90k 0.485 5.25 23.5 24.2 LINK
MaskRCNN-R50-FPN-RFS Cosine 640-800 90k 0.485 5.25 24.5 24.9 LINK
MaskRCNN-R50-FPN-DisAlign 640-800 90k 0.486 5.26 23.7 24.3 LINK
MaskRCNN-R50-FPN-DisAlign Cosine 640-800 90k 0.500 5.26 26.3 27.1 LINK
MaskRCNN-R50-FPN-RFS-DisAlign Cosine 640-800 90k 0.500 5.26 27.1 27.5 LINK

ResNet-101

Name Cls Norm input size lr sched train time (s/iter) train mem (GB) box AP mask AP Trained Model
MaskRCNN-R101-FPN 640-800 90k 22.6 22.8 LINK
MaskRCNN-R101-FPN Cosine 640-800 90k 24.8 25.3 LINK
MaskRCNN-R101-FPN-RFS Cosine 640-800 90k 26.6 26.8 LINK
MaskRCNN-R101-FPN-DisAlign 640-800 90k 25.9 26.2 LINK
MaskRCNN-R101-FPN-DisAlign Cosine 640-800 90k 27.6 28.1 LINK
MaskRCNN-R101-FPN-RFS-DisAlign Cosine 640-800 90k 28.7 28.9 LINK

ResNeXt-101

Name Cls Norm input size lr sched train time (s/iter) train mem (GB) box AP mask AP Trained Model
MaskRCNN-X101-FPN 640-800 90k 24.8 25.2 LINK
MaskRCNN-X101-FPN Cosine 640-800 90k 27.4 28.4 LINK
MaskRCNN-X101-FPN-DisAlign 640-800 90k 26.9 27.3 LINK
MaskRCNN-X101-FPN-DisAlign Cosine 640-800 90k 29.6 30.2 LINK

Cascade R-CNN

LVIS 1.0(Mask R-CNN)

ResNet-50

Name Cls Norm input size lr sched train time (s/iter) train mem (GB) box AP mask AP Trained Model
MaskRCNN-R50-FPN 640-800 180k 0.486 5.26 18.8 18.3 LINK
MaskRCNN-R50-FPN Cosine 640-800 180k 0.500 5.26 21.3 21.1 LINK
MaskRCNN-R50-FPN-RFS 640-800 180k 0.485 5.25 22.9 22.5 LINK
MaskRCNN-R50-FPN-RFS(A1) 640-800 180k 22.3
MaskRCNN-R50-FPN-DisAlign 640-800 180k 0.486 5.26 21.9 21.3 LINK
MaskRCNN-R50-FPN-DisAlign Cosine 640-800 180k 0.500 5.26 24.8 24.2 LINK

ResNet-101

ResNeXt-101

  • A1: Evaluating Large-Vocabulary Object Detectors: The Devil is in the Details