Fully Convolutional Networks for Semantic Segmentation

This is the reference implementation of the models and code for the fully convolutional networks (FCNs) in the PAMI FCNand CVPR FCN papers:

Fully Convolutional Models for Semantic Segmentation
Evan Shelhamer*, Jonathan Long*, Trevor Darrell
PAMI 2016
arXiv:1605.06211

Fully Convolutional Models for Semantic Segmentation
Jonathan Long*, Evan Shelhamer*, Trevor Darrell
CVPR 2015
arXiv:1411.4038

Note that this is a work in progress and the final, reference version is coming soon.

Please ask Caffe and FCN usage questions on the caffe-users mailing list.

Refer to these slides for a summary of the approach.

These models are compatible with BVLC/caffe:master. Compatibility has held since master@8c66fa5 with the merge of PRs #3613 and #3570. The code and models here are available under the same license as Caffe (BSD-2) and the Caffe-bundled models (that is, unrestricted use; see the BVLC model license).

PASCAL VOC models:

trained online with high momentum for a ~5 point boost in mean intersection-over-union over the original models. These models are trained using extra data from Hariharan et al., but excluding SBD val. FCN-32s is fine-tuned from the ILSVRC-trained VGG-16 model, and the finer strides are then fine-tuned in turn. The "at-once" FCN-8s is fine-tuned from VGG-16 all-at-once by scaling the skip connections to better condition optimization.

FCN-32s PASCAL: single stream, 32 pixel prediction stride net, scoring 63.6 mIU on seg11valid

FCN-16s PASCAL: two stream, 16 pixel prediction stride net, scoring 65.0 mIU on seg11valid

FCN-8s PASCAL: three stream, 8 pixel prediction stride net, scoring 65.5 mIU on seg11valid and 67.2 mIU on seg12test

FCN-8s PASCAL at-once: all-at-once, three stream, 8 pixel prediction stride net, scoring 65.4 mIU on seg11valid

FCN-AlexNet PASCAL: AlexNet (CaffeNet) architecture, single stream, 32 pixel prediction stride net, scoring 48.0 mIU on seg11valid. Unlike the FCN-32/16/8s models, this network is trained with gradient accumulation, normalized loss, and standard momentum. (Note: when both FCN-32s/FCN-VGG16 and FCN-AlexNet are trained in this same way FCN-VGG16 is far better; see Table 1 of the paper.)

To reproduce the validation scores, use the seg11valid split defined by the paper in footnote 7. Since SBD train and PASCAL VOC 2011 segval intersect, we only evaluate on the non-intersecting set for validation purposes.

NYUDv2 models:

trained online with high momentum on color, depth, and HHA features (from Gupta et al. https://github.com/s-gupta/rcnn-depth). These models demonstrate FCNs for multi-modal input.

FCN-32s NYUDv2 Color: single stream, 32 pixel prediction stride net on color/BGR input

FCN-32s NYUDv2 HHA: single stream, 32 pixel prediction stride net on HHA input

FCN-32s NYUDv2 Early Color-Depth: single stream, 32 pixel prediction stride net on early fusion of color and (log) depth for 4-channel input

FCN-32s NYUDv2 Late Color-HHA: single stream, 32 pixel prediction stride net by late fusion of FCN-32s NYUDv2 Color and FCN-32s NYUDv2 HHA

SIFT Flow models:

trained online with high momentum for joint semantic class and geometric class segmentation. These models demonstrate FCNs for multi-task output.

FCN-32s SIFT Flow: single stream stream, 32 pixel prediction stride net

FCN-16s SIFT Flow: two stream, 16 pixel prediction stride net

FCN-8s SIFT Flow: three stream, 8 pixel prediction stride net

Frequently Asked Questions

Is learning the interpolation necessary?

In our original experiments the interpolation layers were initialized to bilinear kernels and then learned. In follow-up experiments, and this reference implementation, the bilinear kernels are fixed. There is no significant difference in accuracy in our experiments, and fixing these parameters gives a slight speed-up. Note that in our networks there is only one interpolation kernel per output class, and results may differ for higher-dimensional and non-linear interpolation, for which learning may help further.

Why pad the input?

The 100 pixel input padding guarantees that the network output can be aligned to the input for any input size in the given datasets, for instance PASCAL VOC. The alignment is handled automatically by net specification and the crop layer. It is possible, though less convenient, to calculate the exact offsets necessary and do away with this amount of padding.

Why are all the outputs/gradients/parameters zero?

This is almost universally due to not initializing the weights as needed. To reproduce our FCN training, or train your own FCNs, it is crucial to transplant the weights from the corresponding ILSVRC net such as VGG16. The included surgery.transplant() method can help with this.

What about FCN-GoogLeNet?

a reference FCN-GoogLeNet for PASCAL VOC is coming soon.