CornerNet

open-mmlab / mmdetection

Last updated on Feb 23, 2021
CornerNet (HourglassNet-104, bs=10 x 5)
Memory (M) 13900.0
inference time (s/im) 0.2381
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Convolution, Corner Pooling, Stacked Hourglass Network
Batch Size 10 x 5
Memory (M) 13900.0
inference time (s/im) 0.2381
SHOW MORE
SHOW LESS
CornerNet (HourglassNet-104, bs=32 x 3)
Memory (M) 9500.0
inference time (s/im) 0.25641
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Convolution, Corner Pooling, Stacked Hourglass Network
Batch Size 32 x 3
Memory (M) 9500.0
inference time (s/im) 0.25641
SHOW MORE
SHOW LESS
CornerNet (HourglassNet-104, bs=8 x 6)
Memory (M) 15900.0
inference time (s/im) 0.2381
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Convolution, Corner Pooling, Stacked Hourglass Network
Batch Size 8 x 6
Memory (M) 15900.0
inference time (s/im) 0.2381
SHOW MORE
SHOW LESS
README.md

CornerNet

Introduction

[ALGORITHM]

@inproceedings{law2018cornernet,
  title={Cornernet: Detecting objects as paired keypoints},
  author={Law, Hei and Deng, Jia},
  booktitle={15th European Conference on Computer Vision, ECCV 2018},
  pages={765--781},
  year={2018},
  organization={Springer Verlag}
}

Results and models

Backbone Batch Size Step/Total Epochs Mem (GB) Inf time (fps) box AP Config Download
HourglassNet-104 10 x 5 180/210 13.9 4.2 41.2 config model | log
HourglassNet-104 8 x 6 180/210 15.9 4.2 41.2 config model | log
HourglassNet-104 32 x 3 180/210 9.5 3.9 40.4 config model | log

Note:

  • TTA setting is single-scale and flip=True.
  • Experiments with images_per_gpu=6 are conducted on Tesla V100-SXM2-32GB, images_per_gpu=3 are conducted on GeForce GTX 1080 Ti.
  • Here are the descriptions of each experiment setting:
  • 10 x 5: 10 GPUs with 5 images per gpu. This is the same setting as that reported in the original paper.
  • 8 x 6: 8 GPUs with 6 images per gpu. The total batchsize is similar to paper and only need 1 node to train.
  • 32 x 3: 32 GPUs with 3 images per gpu. The default setting for 1080TI and need 4 nodes to train.

Results

Object Detection on COCO minival

Object Detection
BENCHMARK MODEL METRIC NAME METRIC VALUE GLOBAL RANK
COCO minival CornerNet (HourglassNet-104, bs=10 x 5) box AP 41.2 # 63
COCO minival CornerNet (HourglassNet-104, bs=8 x 6) box AP 41.2 # 63
COCO minival CornerNet (HourglassNet-104, bs=32 x 3) box AP 40.4 # 70
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