This is an (re-)implementation of DeepLab-ResNet in TensorFlow for semantic image segmentation on the PASCAL VOC dataset.
If you encounter some problems and would like to create an issue, please read this first. If the guide below does not cover your question, please use search to see if a similar issue has already been solved before. Finally, if you are unable to find an answer, please fill in the issue with details of your problem provided.
All the experiments are been done using python2.7
. python3
will likely require some minor modifications.
After training, I have multiple files that look like model.ckpt-xxxx.index
, model.ckpt-xxxx.dataxxxx
and model.ckpt-xxxx.meta
. Which one of them should I use to restore the model for inference?
Instead of providing a path to one of those files, you must provide just model.ckpt-xxxx
. It will fetch other files.
First, check that your images are being read correctly. The setup implies that segmentation masks are saved without a colour map, i.e., each pixel contains a class index, not an RGB value. Second, tune your hyperparameters. As there are no general strategies that work for each case, the design of this procedure is up to you.
Please refer to this topic.
29 Jan, 2017:
- Fixed the implementation of the batch normalisation layer: it now supports both the training and inference steps. If the flag
--is-training
is provided, the running means and variances will be updated; otherwise, they will be kept intact. The.ckpt
files have been updated accordingly - to download please refer to the new link provided below. - Image summaries during the training process can now be seen using TensorBoard.
- Fixed the evaluation procedure: the 'void' label (
255
) is now correctly ignored. As a result, the performance score on the validation set has increased to80.1%
.
11 Feb, 2017:
- The training script
train.py
has been re-written following the original optimisation setup: SGD with momentum, weight decay, learning rate with polynomial decay, different learning rates for different layers, ignoring the 'void' label (255
). - The training script with multi-scale inputs
train_msc.py
has been added: the input is resized to0.5
and0.75
of the original resolution, and4
losses are aggregated: loss on the original resolution, on the0.75
resolution, on the0.5
resolution, and loss on the all fused outputs. - Evaluation of a single-scale converted pre-trained model on the PASCAL VOC validation dataset (using 'SegmentationClassAug') leads to
86.9%
mIoU (as trainval was likely to be used for final training). This is confirmed by the official PASCAL VOC server. The score on the test dataset is75.8%
.
22 Feb, 2017:
- The training script with multi-scale inputs
train_msc.py
now supports gradients accumulation: the relevant parameter--grad-update-every
effectively mimics the behaviour ofiter_size
of Caffe. This allows to use batches of bigger sizes with less GPU memory being consumed. (Thanks to @arslan-chaudhry for this contribution!) - The random mirror and random crop options have been added. (Again big thanks to @arslan-chaudhry !)
23 Apr, 2017:
- TensorFlow 1.1.0 is now supported.
- Three new flags
--num-classes
,--ignore-label
and--not-restore-last
are added to ease the usability of the scripts on new datasets. Check out these instructions on how to set up the training process on your dataset.
The DeepLab-ResNet is built on a fully convolutional variant of ResNet-101 with atrous (dilated) convolutions, atrous spatial pyramid pooling, and multi-scale inputs (not implemented here).
The model is trained on a mini-batch of images and corresponding ground truth masks with the softmax classifier at the top. During training, the masks are downsampled to match the size of the output from the network; during inference, to acquire the output of the same size as the input, bilinear upsampling is applied. The final segmentation mask is computed using argmax over the logits.
Optionally, a fully-connected probabilistic graphical model, namely, CRF, can be applied to refine the final predictions.
On the test set of PASCAL VOC, the model achieves 79.7%
of mean intersection-over-union.
For more details on the underlying model please refer to the following paper:
@article{CP2016Deeplab,
title={DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs},
author={Liang-Chieh Chen and George Papandreou and Iasonas Kokkinos and Kevin Murphy and Alan L Yuille},
journal={arXiv:1606.00915},
year={2016}
}
TensorFlow needs to be installed before running the scripts. TensorFlow v1.1.0 is supported; for TensorFlow v0.12 please refer to this branch; for TensorFlow v0.11 please refer to this branch. Note that those branches may not have the same functional as the current master.
To install the required python packages (except TensorFlow), run
pip install -r requirements.txt
or for a local installation
pip install --user -r requirements.txt
To imitate the structure of the model, we have used .caffemodel
files provided by the authors. The conversion has been performed using Caffe to TensorFlow with an additional configuration for atrous convolution and batch normalisation (since the batch normalisation provided by Caffe-tensorflow only supports inference).
There is no need to perform the conversion yourself as you can download the already converted models - deeplab_resnet.ckpt
(pre-trained) and deeplab_resnet_init.ckpt
(the last layers are randomly initialised) - here.
Nevertheless, it is easy to perform the conversion manually, given that the appropriate .caffemodel
file has been downloaded, and Caffe to TensorFlow dependencies have been installed. The Caffe model definition is provided in misc/deploy.prototxt
.
To extract weights from .caffemodel
, run the following:
python convert.py /path/to/deploy/prototxt --caffemodel /path/to/caffemodel --data-output-path /where/to/save/numpy/weights
As a result of running the command above, the model weights will be stored in /where/to/save/numpy/weights
. To convert them to the native TensorFlow format (.ckpt
), simply execute:
python npy2ckpt.py /where/to/save/numpy/weights --save-dir=/where/to/save/ckpt/weights
To train the network, one can use the augmented PASCAL VOC 2012 dataset with 10582
images for training and 1449
images for validation.
The training script allows to monitor the progress in the optimisation process using TensorBoard's image summary. Besides that, one can also exploit random scaling and mirroring of the inputs during training as a means for data augmentation. For example, to train the model from scratch with random scale and mirroring turned on, simply run:
python train.py --random-mirror --random-scale
To see the documentation on each of the training settings run the following:
python train.py --help
An additional script, fine_tune.py
, demonstrates how to train only the last layers of the network. The script train_msc.py
with multi-scale inputs fully resembles the training setup of the original model.
The single-scale model shows 86.9%
mIoU on the Pascal VOC 2012 validation dataset ('SegmentationClassAug'). No post-processing step with CRF is applied.
The following command provides the description of each of the evaluation settings:
python evaluate.py --help
To perform inference over your own images, use the following command:
python inference.py /path/to/your/image /path/to/ckpt/file
This will run the forward pass and save the resulted mask with this colour map:
In order to apply the same scripts using your own dataset, you would need to follow the next steps:
- Make sure that your segmentation masks are in the same format as the ones in the DeepLab setup (i.e., without a colour map). This means that if your segmentation masks are RGB images, you would need to convert each 3-D RGB vector into a 1-D label. For example, take a look here;
- Create a file with instances of your dataset in the same format as in files here;
- Change the flags
data-dir
anddata-list
accordingly in thehttps://gist.github.com/DrSleep/4bce37254c5900545e6b65f6a0858b9c); script file that you will be using (e.g.,python train.py --data-dir /my/data/dir --data-list /my/data/list
); - Change the
IMG_MEAN
vector accordingly in the script file that you will be using; - For visualisation purposes, you will also need to change the colour map here;
- Change the flags
num-classes
andignore-label
accordingly in the script that you will be using (e.g.,python train.py --ignore-label 255 --num-classes 21
). - If restoring weights from the
PASCAL
models for your dataset with a different number of classes, you will also need to pass the--not-restore-last
flag, which will prevent the last layers of size21
from being restored.
The post-processing step with CRF is currently being implemented here.