Max Iter	90000
lr sched	1x
Backbone Layers	50

Max Iter	540000
lr sched	6x
Backbone Layers	50

TensorMask

facebookresearch / detectron2

Last updated on Feb 19, 2021

Parameters 36 Million

FLOPs 504 Billion

File Size 139.74 MB

Training Data MS COCO

Training Resources 8 NVIDIA V100 GPUs

Training Time

Architecture	ResNet, FPN, Focal Loss
Max Iter	90000
lr sched	1x
Backbone Layers	50
SHOW MORE
SHOW LESS

Parameters 36 Million

FLOPs 504 Billion

File Size 139.74 MB

Training Data MS COCO

Training Resources 8 NVIDIA V100 GPUs

Training Time

Architecture	ResNet, FPN, Focal Loss
Max Iter	540000
lr sched	6x
Backbone Layers	50
SHOW MORE
SHOW LESS

README.md

Summary

TensorMask is a method for dense object segmentation which treats dense instance segmentation as a prediction task over 4D tensors, explicitly capturing this geometry and enabling novel operators on 4D tensors.

Installation

First install Detectron2 following the documentation and setup the dataset. Then compile the TensorMask-specific op (swap_align2nat):

pip install -e /path/to/detectron2/projects/TensorMask

Training

To train a model, run:

python /path/to/detectron2/projects/TensorMask/train_net.py --config-file <config.yaml>

For example, to launch TensorMask BiPyramid training (1x schedule) with ResNet-50 backbone on 8 GPUs, one should execute:

python /path/to/detectron2/projects/TensorMask/train_net.py --config-file configs/tensormask_R_50_FPN_1x.yaml --num-gpus 8

Evaluation

Model evaluation can be done similarly (6x schedule with scale augmentation):

python /path/to/detectron2/projects/TensorMask/train_net.py --config-file configs/tensormask_R_50_FPN_6x.yaml --eval-only MODEL.WEIGHTS /path/to/model_checkpoint

Citation

@InProceedings{chen2019tensormask,
  title={Tensormask: A Foundation for Dense Object Segmentation},
  author={Chen, Xinlei and Girshick, Ross and He, Kaiming and Doll{\'a}r, Piotr},
  journal={The International Conference on Computer Vision (ICCV)},
  year={2019}
}

Results

Object Detection on COCO minival

Object Detection

BENCHMARK	MODEL	METRIC NAME	METRIC VALUE	GLOBAL RANK
COCO minival	TensorMask (R50, 6x)	box AP	41.4	# 61
COCO minival	TensorMask (R50, 1x)	box AP	37.6	# 98

Instance Segmentation

BENCHMARK	MODEL	METRIC NAME	METRIC VALUE	GLOBAL RANK
COCO minival	TensorMask (R50, 6x)	mask AP	35.8	# 48
COCO minival	TensorMask (R50, 1x)	mask AP	32.4	# 55