Counting 3,567 Big Data & Machine Learning Frameworks, Toolsets, and Examples...
Suggestion? Feedback? Tweet @stkim1

Last Commit
Feb. 23, 2019
Jan. 7, 2019

ExtremeNet: Training and Evaluation Code

Code for bottom-up object detection by grouping extreme and center points:

Bottom-up Object Detection by Grouping Extreme and Center Points,
Xingyi Zhou, Jiacheng Zhuo, Philipp Krähenbühl,
arXiv technical report (1901.08043)

This project is developed upon the CornerNet code and contains the code from Deep Extreme Cut. Thanks to the original authors!

Contact: Any questions or discussions are welcomed!


With the advent of deep learning, object detection drifted from a bottom-up to a top-down recognition problem. State of the art algorithms enumerate a near-exhaustive list of object locations and classify each into: object or not. In this paper, we show that bottom-up approaches still perform competitively. We detect four extreme points (top-most, left-most, bottom-most, right-most) and one center point of objects using a standard keypoint estimation network. We group the five keypoints into a bounding box if they are geometrically aligned. Object detection is then a purely appearance-based keypoint estimation problem, without region classification or implicit feature learning. The proposed method performs on-par with the state-of-the-art region based detection methods, with a bounding box AP of 43.2% on COCO test-dev. In addition, our estimated extreme points directly span a coarse octagonal mask, with a COCO Mask AP of 18.9%, much better than the Mask AP of vanilla bounding boxes. Extreme point guided segmentation further improves this to 34.6% Mask AP.


The code was tested with Anaconda Python 3.6 and PyTorch v0.4.1. After install Anaconda:

  1. Clone this repo:

    git clone --recursive ExtremeNet_ROOT
  2. Create an Anaconda environment using the provided package list from Cornernet.

    conda create --name CornerNet --file conda_packagelist.txt
    source activate CornerNet
  3. Compiling NMS (originally from Faster R-CNN and Soft-NMS).

    cd ExtremeNet_ROOT/external


  • Download our pre-trained model and put it in cache/.

  • Optionally, if you want to test instance segmentation with Deep Extreme Cut, download their PASCAL + SBD pertained model and put it in cache/.

  • Run the demo

    python [--demo /path/to/image/or/folder] [--show_mask]

    Contents in [] are optional. By default, it runs the sample images provided in ExtremeNet_ROOT/images/ (from Detectron). We show the predicted extreme point heatmaps (combined four heatmaps and overlaid on the input image), the predicted center point heatmap, and the detection and octagon mask results. If setup correctly, the output will look like:

    If --show_mask is turned on, it further pipelined with Deep Extreme Cut for instance segmentation. The output will look like:

Data preparation

If you want to reproduce the results in the paper for benchmark evaluation and training, you will need to setup dataset.

Installing MS COCO APIs

cd ExtremeNet_ROOT/data
git clone coco
cd ExtremeNet_ROOT/data/coco/PythonAPI
python install --user

Downloading MS COCO Data

  • 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.

  • Place the data (or create symlinks) to make the data folder like:

    |-- data
    `-- |-- coco
        `-- |-- annotations
            |   |-- instances_train2017.json
            |   |-- instances_val2017.json
            |   |-- image_info_test-dev2017.json
            `-- images
                |-- train2017
                |-- val2017
                |-- test2017

Generate extreme point annotation from segmentation:

cd ExtremeNet_ROOT/tools/

It generates instances_extreme_train2017.json and instances_extreme_val2017.json in data/coco/annotations/.

Benchmark Evaluation

After downloading our pre-trained model and the dataset,

  • Run the following command to evaluate object detection:

    python ExtremeNet [--suffix multi_scale]

    The results on COCO validation set should be 40.3 box AP without --suffix multi_scale and 43.3 box AP (download) with --suffix multi_scale.

  • After obtaining the detection results, run the following commands for instance segmentation:

    python results/ExtremeNet/250000/validation/multi_scale/results.json

    The results on COCO validation set should be 34.6 mask AP(The evaluation will be slow).

  • You can test with other hyper-parameters by creating a new config file (ExtremeNet-<suffix>.json) in config/.


You will need 5x 12GB GPUs to reproduce our training. Our model is fine-tuned on the 10-GPU pre-trained CornerNet model. After downloading the CornerNet model and put it in cache/, run

python ExtremeNet

You can resume a half-trained model by

python ExtremeNet --iter xxxx


  • Training takes about 10 days in our Titan V GPUs. Train with 150000 iterations (about 6 days) will be 0.5 AP lower.
  • Training from scratch for the same iteration (250000) may result in 2 AP lower than fintuning from CornerNet, but can get higher performance (43.9AP on COCO val w/ multi-scale testing) if trained for 500000 iterations
  • Changing the focal loss implementation to this can accelerate training, but costs more GPU memory.


If you find this model useful for your research, please use the following BibTeX entry.

  title={Bottom-up Object Detection by Grouping Extreme and Center Points},
  author={Zhou, Xingyi and Zhuo, Jiacheng and Kr{\"a}henb{\"u}hl, Philipp},
  booktitle={arXiv preprint arXiv:1901.08043},

Please also considering citing the CornerNet paper (where this code is heavily borrowed from) and Deep Extreme Cut paper (if you use the instance segmentation part).

  title={CornerNet: Detecting Objects as Paired Keypoints},
  author={Law, Hei and Deng, Jia},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},

  Title          = {Deep Extreme Cut: From Extreme Points to Object Segmentation},
  Author         = {K.K. Maninis and S. Caelles and J. Pont-Tuset and L. {Van Gool}},
  Booktitle      = {Computer Vision and Pattern Recognition (CVPR)},
  Year           = {2018}