The code for implementing the TTFNet.


  • Simple: Anchor-free, single-stage, light-head, no time-consuming post-processing. TTFNet only requires two detection heads for object localization and size regression, respectively.
  • Training Time Friendly: Our TTFNet outperforms a range of real-time detectors while suppressing them in training time. Moreover, super-fast TTFNet-18 and TTFNet-53 can reach 25.9 AP / 112 FPS only after 2 hours and 32.9 AP / 55 FPS after about 3 hours on the MS COCO dataset using 8 GTX 1080Ti.
  • Fast and Precise: Our TTFNet-18/34/53 can achieve 28.1AP / 112FPS, 31.3AP / 87FPS, and 35.1AP / 54 FPS on 1 GTX 1080Ti.




TT stands for training time. * indicates that the model is not presented in the original paper, but we list these results to explore the performances of these work when adopting a light backbone network.

All the training time is measured on 8 GTX 1080Ti, and all the inference speed is measured using converged models on 1 GTX 1080Ti. Note that the training time does not include the time consumed by evaluation.


Our TTFNet is based on mmdetection. Please check for installation instructions, and you may want to see the original We will submit a pull request soon.

Note that the darknet part was transplanted (i.e., MXNet => Pytorch) from another toolbox Gluoncv. In addition, portions of the code are borrowed from CornerNet and CenterNet. Thanks for their work !


We provide the following converged models.

Model Training Hours FPS AP(minival) Link
TTFNet-18 (1x) 1.8 112.2 25.9 Download
TTFNet-18 (2x) 3.6 112.3 28.1 Download
TTFNet-34 (2x) 4.1 86.6 31.3 Download
TTFNet-53 (1x) 3.1 54.8 32.9 Download
TTFNet-53 (2x) 6.1 54.4 35.1 Download

We also provide the pretrained Darknet53 here.

The following command will evaluate converged TTFNet-53 on 8 GPUs:

./tools/ configs/ttfnet/ /path/to/the/checkpoint 8


The following commands will train TTFNet-18 on 8 GPUs for 24 epochs and TTFNet-53 on 8 GPUs for 12 epochs:

./tools/ configs/ttfnet/ 8
./tools/ configs/ttfnet/ 8


Please consider citing our paper in your publications if the project helps your research. BibTeX reference is as follows.

  title   = {Training-Time-Friendly Network for Real-Time Object Detection},
  author  = {Zili Liu, Tu Zheng, Guodong Xu, Zheng Yang, Haifeng Liu, Deng Cai},
  journal = {arXiv preprint arXiv:1909.00700},
  year    = {2019}