HTC

open-mmlab / mmdetection

Last updated on Feb 23, 2021
HTC DCN (X-101-64x4d-FPN, 20e, pytorch, DCN=c3-c5)
lr sched 20e
Backbone Layers 101
File Size 547.55 MB
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture RPN, ResNeXt, Convolution, Deformable Convolution, FPN, 1x1 Convolution, HTC, RoIAlign
lr sched 20e
Backbone Layers 101
SHOW MORE
SHOW LESS
HTC (R-101-FPN, 20e, pytorch)
Memory (M) 10200.0
inference time (s/im) 0.18182
File Size 379.14 MB
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture RPN, HTC, Convolution, FPN, 1x1 Convolution, ResNet, RoIAlign
lr sched 20e
Memory (M) 10200.0
Backbone Layers 101
inference time (s/im) 0.18182
SHOW MORE
SHOW LESS
HTC (R-50-FPN, 1x, pytorch)
Memory (M) 8200.0
inference time (s/im) 0.17241
File Size 306.44 MB
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture RPN, HTC, Convolution, FPN, 1x1 Convolution, ResNet, RoIAlign
lr sched 1x
Memory (M) 8200.0
Backbone Layers 50
inference time (s/im) 0.17241
SHOW MORE
SHOW LESS
HTC (R-50-FPN, 20e, pytorch)
Memory (M) 8200.0
Backbone Layers 50
File Size 306.44 MB
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture RPN, HTC, Convolution, FPN, 1x1 Convolution, ResNet, RoIAlign
lr sched 20e
Memory (M) 8200.0
Backbone Layers 50
SHOW MORE
SHOW LESS
HTC (X-101-32x4d-FPN, 20e, pytorch)
Memory (M) 11400.0
inference time (s/im) 0.2
File Size 377.84 MB
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture RPN, ResNeXt, Convolution, FPN, 1x1 Convolution, HTC, RoIAlign
lr sched 20e
Memory (M) 11400.0
Backbone Layers 101
inference time (s/im) 0.2
SHOW MORE
SHOW LESS
HTC (X-101-64x4d-FPN, 20e, pytorch)
Memory (M) 14500.0
inference time (s/im) 0.22727
File Size 527.92 MB
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture RPN, ResNeXt, Convolution, FPN, 1x1 Convolution, HTC, RoIAlign
lr sched 20e
Memory (M) 14500.0
Backbone Layers 101
inference time (s/im) 0.22727
SHOW MORE
SHOW LESS
README.md

Hybrid Task Cascade for Instance Segmentation

Introduction

[ALGORITHM]

We provide config files to reproduce the results in the CVPR 2019 paper for Hybrid Task Cascade.

@inproceedings{chen2019hybrid,
  title={Hybrid task cascade for instance segmentation},
  author={Chen, Kai and Pang, Jiangmiao and Wang, Jiaqi and Xiong, Yu and Li, Xiaoxiao and Sun, Shuyang and Feng, Wansen and Liu, Ziwei and Shi, Jianping and Ouyang, Wanli and Chen Change Loy and Dahua Lin},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
  year={2019}
}

Dataset

HTC requires COCO and COCO-stuff dataset for training. You need to download and extract it in the COCO dataset path. The directory should be like this.

mmdetection
├── mmdet
├── tools
├── configs
├── data
│   ├── coco
│   │   ├── annotations
│   │   ├── train2017
│   │   ├── val2017
│   │   ├── test2017
|   |   ├── stuffthingmaps

Results and Models

The results on COCO 2017val are shown in the below table. (results on test-dev are usually slightly higher than val)

Backbone Style Lr schd Mem (GB) Inf time (fps) box AP mask AP Config Download
R-50-FPN pytorch 1x 8.2 5.8 42.3 37.4 config model | log
R-50-FPN pytorch 20e 8.2 - 43.3 38.3 config model | log
R-101-FPN pytorch 20e 10.2 5.5 44.8 39.6 config model | log
X-101-32x4d-FPN pytorch 20e 11.4 5.0 46.1 40.5 config model | log
X-101-64x4d-FPN pytorch 20e 14.5 4.4 47.0 41.4 config model | log
  • In the HTC paper and COCO 2018 Challenge, score_thr is set to 0.001 for both baselines and HTC.
  • We use 8 GPUs with 2 images/GPU for R-50 and R-101 models, and 16 GPUs with 1 image/GPU for X-101 models. If you would like to train X-101 HTC with 8 GPUs, you need to change the lr from 0.02 to 0.01.

We also provide a powerful HTC with DCN and multi-scale training model. No testing augmentation is used.

Backbone Style DCN training scales Lr schd box AP mask AP Config Download
X-101-64x4d-FPN pytorch c3-c5 400~1400 20e 50.4 43.8 config model | log

Results

Object Detection on COCO minival

Object Detection on COCO minival
MODEL BOX AP
HTC DCN (X-101-64x4d-FPN, 20e, pytorch, DCN=c3-c5) 50.4
HTC (X-101-64x4d-FPN, 20e, pytorch) 47.0
HTC (X-101-32x4d-FPN, 20e, pytorch) 46.1
HTC (R-101-FPN, 20e, pytorch) 44.8
HTC (R-50-FPN, 20e, pytorch) 43.3
HTC (R-50-FPN, 1x, pytorch) 42.3
Instance Segmentation on COCO minival
MODEL MASK AP
HTC DCN (X-101-64x4d-FPN, 20e, pytorch, DCN=c3-c5) 43.8
HTC (X-101-64x4d-FPN, 20e, pytorch) 41.4
HTC (X-101-32x4d-FPN, 20e, pytorch) 40.5
HTC (R-101-FPN, 20e, pytorch) 39.6
HTC (R-50-FPN, 20e, pytorch) 38.3
HTC (R-50-FPN, 1x, pytorch) 37.4
Contact us on: hello@paperswithcode.com . Papers With Code is a free resource with all data licensed under CC-BY-SA.
Terms Data policy Cookies policy