Skip to content

Latest commit

 

History

History
77 lines (53 loc) · 3.75 KB

GETTING_STARTED.md

File metadata and controls

77 lines (53 loc) · 3.75 KB
Raw

Getting Started

This document provides information about how to train and evaluate HoughNet on COCO. First, make sure that you have completed the installation.

Dataset preparation

  • Download the images (2017 Train, 2017 Val, 2017 Test) from coco website.

  • Download annotation files (2017 train/val and test image info) from coco website.

  • If you like to use minicoco dataset, download the json file and place it under the annotations folder.

    ${COCO_PATH}
|-- annotations
    |-- instances_train2017.json
    |-- instances_minitrain2017.json
    |-- instances_val2017.json
    |-- image_info_test-dev2017.json
|-- images
    |-- train2017
    |-- val2017
    |-- test2017

Evaluation

Download the models you want to evaluate from our model zoo and put them in HoughNet_ROOT/models/.

To evaluate object detection with Resnet-101 w DCN on val2017 run

python src/test.py ctdet --houghnet --exp_id coco_resdcn_101_light --arch resdcn_101 --keep_res --resume  --load_model ./models/ctdet_coco_resdcn101_light.pth --coco_path $COCO_PATH

This will give an AP of 35.7 on val2017. --keep_res is for keeping the original image resolution. Without --keep_res it will resize the images to 512 x 512. You can add --flip_test and --flip_test --test_scales 0.6,0.8,1,1.2,1.5,1.8 to the above command, for flip test and multi-scale test, respectively. The expected APs on val2017 are 37.2 and 41.5, respectively.

For multi-scale test with Hourglass net, run

python src/test.py ctdet --houghnet --exp_id coco_hg_scratch --arch hourglass --keep_res --resume --flip_test --test_scales 0.6,0.8,1,1.2,1.5,1.8 --load_model ./models/ctdet_coco_resdcn101_light.pth --coco_path $COCO_PATH

More results could be found in the model zoo.

Training

You could find all the training scripts in the experiments folder. In the case that you don't have 4 GPUs, you can follow the linear learning rate rule to adjust the learning rate. For instance, to train COCO object detection with Resnet-101 w DCN model using 4 Tesla V100 GPUs on train2017, run

python src/main.py ctdet --houghnet --exp_id coco_resdcn_101 --arch resdcn_101 --batch_size 44 --master_batch 8 --lr 1.75e-4 --gpus 0,1,2,3 --num_workers 16 --coco_path $COCO_PATH

or on minitrain

python src/main.py ctdet --houghnet --minicoco --exp_id coco_resdcn_101 --arch resdcn_101 --batch_size 44 --master_batch 8 --lr 1.75e-4 --gpus 0,1,2,3 --num_workers 16 --coco_path $COCO_PATH

If the training is terminated before finishing, you can use the same command with --resume to resume training. It will find the latest model with the same exp_id.

Our best Hourglass model is finetuned from the pretrained ExtremeNet model (from the ExtremeNet repo). You need to download and load the model for training (see the script).

We also have another Hourglass model finetuned from the pretrained CornerNet model. You could download the model and load the model for training (see the script).

For the training and evaluation of instance segmentation, human keypoint detection and 3D object detection tasks please use their corresponding scripts.