This is my implementation of Deeplab-LargeFOV.
Run the script to download the dataset:
$ sh get_dataset.sh
This will download and extract the VOC2012 dataset together with the augumented VOC dataset. In the paper, the author used the aug version dataset to train the model and then test on the standard VOC2012 val set.
Just run the training script and evaluating script:
$ python3 train.py --cfg config/pascal_voc_2012_multi_scale.py
$ python3 evaluate.py --cfg config/pascal_voc_2012_multi_scale.py
Then you will see the mIOU result as reported by the authors(57.25 mIOU for deeplab-largeFOV).
If you want to see the result reported by the authors, use the configuration file pascal_voc_aug_multi_scale.py:
$ python3 train.py --cfg config/pascal_voc_aug_multi_scale.py
$ python3 eval.py --cfg config/pascal_voc_aug_multi_scale.py
This will train on the augmented dataset, which is also what the authors used in the paper.
The script infer.py is an example how we could use the trained model to implement segmentation on the picture of example.jpg. Just run it to see the performance of this model:
$ python3 infer.py
And you will see the result picture.
- The authors claimed to use the weight pretrained on imagenet to initialize the model, and they provided the initial model which is exported from the framework of caffe. So far, I found no easy way to convert caffe pretrained weights to that of pytorch, so I employ the following tricks as a remedy:
-
Initialize weights of the common parts of the 'deeplab vgg' and 'pytorch standard vgg' with the weights from pytorch model zoo.
-
Initialize the weights of the extraly appended layers with the msra normal distribution random numbers.
-
Implement warmup with the linear strategy mentioned in this paper
-
Upsample the output logits of the model instead of downsample the ground truth as does in deeplabv3.
-
Use the exponential lr scheduler as does in the deeplabv3.
-
Do multi-scale training and multi-scale testing(with flip). The images are cropped to be (497, 497). The training scales are: [0.75, 1, 1.25, 1.5, 1.75, 2.], and the testing scales are: [0.5, 0.75, 1, 1.25, 1.5, 1.75].
-
Since we do not expect the crop size of the inference is too far away from what we use in the training process, I use crop evaluation.
-
the images are also augmented with random variance of brightness, contrast and saturation.
-
Considering the imbalance between the amount of different pixels, I used the on-line hard example mining loss to train the model.
With these tricks, my model achieves 68.72 mIOU(without crf), modestly better than the result reported by the authors.
I also tried to use mixup in hopes of some further boost to the performance, but the result just gave a big slap on my face. I left the mixup code in the program and if you have special insights on this trick, please feel free to add a pull request or open an issue. Many thanks !