Permalink
186 lines (139 sloc) 13.4 KB

TensorFlow DeepLab Model Zoo

We provide deeplab models pretrained several datasets, including (1) PASCAL VOC 2012, (2) Cityscapes, and (3) ADE20K for reproducing our results, as well as some checkpoints that are only pretrained on ImageNet for training your own models.

DeepLab models trained on PASCAL VOC 2012

Un-tar'ed directory includes:

  • a frozen inference graph (frozen_inference_graph.pb). All frozen inference graphs use output stride of 8 and a single eval scale of 1.0. No left-right flips are used, and MobileNet-v2 based models do not include the decoder module.

  • a checkpoint (model.ckpt.data-00000-of-00001, model.ckpt.index)

Model details

We provide several checkpoints that have been pretrained on VOC 2012 train_aug set or train_aug + trainval set. In the former case, one could train their model with smaller batch size and freeze batch normalization when limited GPU memory is available, since we have already fine-tuned the batch normalization for you. In the latter case, one could directly evaluate the checkpoints on VOC 2012 test set or use this checkpoint for demo. Note MobileNet-v2 based models do not employ ASPP and decoder modules for fast computation.

Checkpoint name Network backbone Pretrained dataset ASPP Decoder
mobilenetv2_dm05_coco_voc_trainaug MobileNet-v2
Depth-Multiplier = 0.5
MS-COCO
VOC 2012 train_aug set
N/A N/A
mobilenetv2_dm05_coco_voc_trainval MobileNet-v2
Depth-Multiplier = 0.5
MS-COCO
VOC 2012 train_aug + trainval sets
N/A N/A
mobilenetv2_coco_voc_trainaug MobileNet-v2 MS-COCO
VOC 2012 train_aug set
N/A N/A
mobilenetv2_coco_voc_trainval MobileNet-v2 MS-COCO
VOC 2012 train_aug + trainval sets
N/A N/A
xception65_coco_voc_trainaug Xception_65 MS-COCO
VOC 2012 train_aug set
[6,12,18] for OS=16
[12,24,36] for OS=8
OS = 4
xception65_coco_voc_trainval Xception_65 MS-COCO
VOC 2012 train_aug + trainval sets
[6,12,18] for OS=16
[12,24,36] for OS=8
OS = 4

In the table, OS denotes output stride.

Checkpoint name Eval OS Eval scales Left-right Flip Multiply-Adds Runtime (sec) PASCAL mIOU File Size
mobilenetv2_dm05_coco_voc_trainaug 16 [1.0] No 0.88B - 70.19% (val) 7.6MB
mobilenetv2_dm05_coco_voc_trainval 8 [1.0] No 2.84B - 71.83% (test) 7.6MB
mobilenetv2_coco_voc_trainaug 16
8
[1.0]
[0.5:0.25:1.75]
No
Yes
2.75B
152.59B
0.1
26.9
75.32% (val)
77.33 (val)
23MB
mobilenetv2_coco_voc_trainval 8 [0.5:0.25:1.75] Yes 152.59B 26.9 80.25% (test) 23MB
xception65_coco_voc_trainaug 16
8
[1.0]
[0.5:0.25:1.75]
No
Yes
54.17B
3055.35B
0.7
223.2
82.20% (val)
83.58% (val)
439MB
xception65_coco_voc_trainval 8 [0.5:0.25:1.75] Yes 3055.35B 223.2 87.80% (test) 439MB

In the table, we report both computation complexity (in terms of Multiply-Adds and CPU Runtime) and segmentation performance (in terms of mIOU) on the PASCAL VOC val or test set. The reported runtime is calculated by tfprof on a workstation with CPU E5-1650 v3 @ 3.50GHz and 32GB memory. Note that applying multi-scale inputs and left-right flips increases the segmentation performance but also significantly increases the computation and thus may not be suitable for real-time applications.

DeepLab models trained on Cityscapes

Model details

We provide several checkpoints that have been pretrained on Cityscapes train_fine set. Note MobileNet-v2 based model has been pretrained on MS-COCO dataset and does not employ ASPP and decoder modules for fast computation.

Checkpoint name Network backbone Pretrained dataset ASPP Decoder
mobilenetv2_coco_cityscapes_trainfine MobileNet-v2 MS-COCO
Cityscapes train_fine set
N/A N/A
xception65_cityscapes_trainfine Xception_65 ImageNet
Cityscapes train_fine set
[6, 12, 18] for OS=16
[12, 24, 36] for OS=8
OS = 4
xception71_dpc_cityscapes_trainfine Xception_71 ImageNet
MS-COCO
Cityscapes train_fine set
Dense Prediction Cell OS = 4
xception71_dpc_cityscapes_trainval Xception_71 ImageNet
MS-COCO
Cityscapes trainval_fine and coarse set
Dense Prediction Cell OS = 4

In the table, OS denotes output stride.

Checkpoint name Eval OS Eval scales Left-right Flip Multiply-Adds Runtime (sec) Cityscapes mIOU File Size
mobilenetv2_coco_cityscapes_trainfine 16
8
[1.0]
[0.75:0.25:1.25]
No
Yes
21.27B
433.24B
0.8
51.12
70.71% (val)
73.57% (val)
23MB
xception65_cityscapes_trainfine 16
8
[1.0]
[0.75:0.25:1.25]
No
Yes
418.64B
8677.92B
5.0
422.8
78.79% (val)
80.42% (val)
439MB
xception71_dpc_cityscapes_trainfine 16 [1.0] No 502.07B - 80.31% (val) 445MB
xception71_dpc_cityscapes_trainval 8 [0.75:0.25:2] Yes - - 82.66% (test) 446MB

DeepLab models trained on ADE20K

Model details

We provide some checkpoints that have been pretrained on ADE20K training set. Note that the model has only been pretrained on ImageNet, following the dataset rule.

Checkpoint name Network backbone Pretrained dataset ASPP Decoder
xception65_ade20k_train Xception_65 ImageNet
ADE20K training set
[6, 12, 18] for OS=16
[12, 24, 36] for OS=8
OS = 4
Checkpoint name Eval OS Eval scales Left-right Flip mIOU Pixel-wise Accuracy File Size
xception65_ade20k_train 8 [0.5:0.25:1.75] Yes 45.65% (val) 82.52% (val) 439MB

Checkpoints pretrained on ImageNet

Un-tar'ed directory includes:

  • model checkpoint (model.ckpt.data-00000-of-00001, model.ckpt.index).

Model details

We also provide some checkpoints that are pretrained on ImageNet and/or COCO (as post-fixed in the model name) so that one could use this for training your own models.

  • mobilenet_v2: We refer the interested users to the TensorFlow open source MobileNet-V2 for details.

  • xception_{41,65,71}: We adapt the original Xception model to the task of semantic segmentation with the following changes: (1) more layers, (2) all max pooling operations are replaced by strided (atrous) separable convolutions, and (3) extra batch-norm and ReLU after each 3x3 depthwise convolution are added. We provide three Xception model variants with different network depths.

  • resnet_v1_{50,101}_beta: We modify the original ResNet-101 [10], similar to PSPNet [11] by replacing the first 7x7 convolution with three 3x3 convolutions. See resnet_v1_beta.py for more details.

Model name File Size
xception_41_imagenet 288MB
xception_65_imagenet 447MB
xception_65_imagenet_coco 292MB
xception_71_imagenet 474MB
resnet_v1_50_beta_imagenet 274MB
resnet_v1_101_beta_imagenet 477MB

References

  1. Mobilenets: Efficient convolutional neural networks for mobile vision applications
    Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, Hartwig Adam
    [link]. arXiv:1704.04861, 2017.

  2. Inverted Residuals and Linear Bottlenecks: Mobile Networks for Classification, Detection and Segmentation
    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, Liang-Chieh Chen
    [link]. arXiv:1801.04381, 2018.

  3. Xception: Deep Learning with Depthwise Separable Convolutions
    François Chollet
    [link]. In the Proc. of CVPR, 2017.

  4. Deformable Convolutional Networks -- COCO Detection and Segmentation Challenge 2017 Entry
    Haozhi Qi, Zheng Zhang, Bin Xiao, Han Hu, Bowen Cheng, Yichen Wei, Jifeng Dai
    [link]. ICCV COCO Challenge Workshop, 2017.

  5. The Pascal Visual Object Classes Challenge: A Retrospective
    Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I. Williams, John M. Winn, Andrew Zisserman
    [link]. IJCV, 2014.

  6. Semantic Contours from Inverse Detectors
    Bharath Hariharan, Pablo Arbelaez, Lubomir Bourdev, Subhransu Maji, Jitendra Malik
    [link]. In the Proc. of ICCV, 2011.

  7. The Cityscapes Dataset for Semantic Urban Scene Understanding
    Cordts, Marius, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, Bernt Schiele.
    [link]. In the Proc. of CVPR, 2016.

  8. Microsoft COCO: Common Objects in Context
    Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, Piotr Dollar
    [link]. In the Proc. of ECCV, 2014.

  9. ImageNet Large Scale Visual Recognition Challenge
    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, Li Fei-Fei
    [link]. IJCV, 2015.

  10. Deep Residual Learning for Image Recognition
    Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
    [link]. CVPR, 2016.

  11. Pyramid Scene Parsing Network
    Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, Jiaya Jia
    [link]. In CVPR, 2017.

  12. Scene Parsing through ADE20K Dataset
    Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, Antonio Torralba
    [link]. In CVPR, 2017.