Skip to content

Latest commit

 

History

History
58 lines (42 loc) · 5.19 KB

MODEL_ZOO.md

File metadata and controls

58 lines (42 loc) · 5.19 KB
Raw

MODEL ZOO

Notes

  • All models are trained on a cluster with 4 Tesla V100 GPUs.
  • The experiments are run with pytorch 1.4.0, CUDA 10.0, and CUDNN 7.5.
  • Testing times are measured on a machine with a single Tesla V100 GPU.
  • All models are trained on COCO train2017 and evaluated on val2017.
  • You could test on COCO test-dev adding --trainval.
  • The models can be downloaded directly from Google drive.

COCO Object Detection Results

Model Test time (ms) AP Download
ctdet_coco_res101 164 / 290 / 2462 34.3 / 36.0 / 40.7 model
ctdet_coco_resdcn101_light 55 / 73 / 498 35.7 / 37.2 / 41.5 model
ctdet_coco_resdcn101 167 / 302 / 1590 35.9 / 37.3 / 41.6 model
ctdet_coco_hg104_scratch 111 / 168 / 1326 39.3 / 40.9 / 43.7 model
ctdet_coco_hg104_cornernet 112 / 172 / 1336 39.8 / 41.7 / 44.3 model
ctdet_coco_hg104_extremenet 117 / 178 / 1328 41.1 / 43.0 / 46.1 model
  • We show test time and AP with no augmentation / flip augmentation / multi scale (0.6, 0.8, 1, 1.2, 1.5, 1.8) augmentation.
  • Testing time includes network forwarding time, decoding time, and nms time (for MS test).

COCO Instance Segmentation Results

Model AP / AP50 Box AP / AP50 Download
ctseg_coco_resdcn101_baseline 27.2 / 46.4 33.9 / 51.3 model
ctseg_coco_resdcn101_light 28.4 / 48.0 35.0 / 52.9 model
  • Results are obtained without any test time augmentation.
  • For instance segmentation task we extended the model with two new branches: prototype mask prediction branch and attention map prediction branch. More information could be found in the paper.

COCO 2D Keypoint Estimation Results

Model AP / AP50 Box AP / AP50 Download
Baseline (CenterNet) 54.7 / 81.7 47.5 / 63.9 model
multi_pose_hm_coco_dla34_1x_light 56.9 / 81.6 50.1 / 71.4 model
multi_pose_hp_coco_dla34_1x_light 56.8 / 81.5 50.2 / 70.9 model
multi_pose_hm_hp_coco_dla34_1x_light 56.9 / 81.6 50.4 / 71.7 model
  • Results are presented with test time flip augmentation.
  • For fair comparison, following CenterNet we fine-tuned the models from their corresponding object detection model.

KITTI 3D Object Detection

Model APe APm APh AOSe AOSm AOSh BEV APe BEV APm BEV APh Download
ddd_coco_dla34_light 89.4 79.2 69.7 85.7 75.1 65.7 35.3 29.7 24.6 model
  • For KITTI dataset please follow the instructions here
  • Experimented with the split from is SubCNN.
  • Results are obtained without any test time augmentation.