FCNs are a model for real-time neural network for class-wise image segmentation. As the name implies, every weight layer in the network is convolutional. The final layer has the same height/width as the input image, making FCNs a useful tool for doing dense pixel-wise predictions without a significant amount of postprocessing. Being fully convolutional also provides great flexibility in the resolutions this model can handle.
This specific model detects 20 different classes. The models have been pre-trained on the COCO train2017 dataset on this class subset.
| Model | Download | Download (with sample test data) | ONNX version | Opset version | Mean IoU |
|---|---|---|---|---|---|
| FCN ResNet-50 | 134 MB | 213 MB | 1.8.0 | 11 | 60.5% |
| FCN ResNet-101 | 207 MB | 281 MB | 1.8.0 | 11 | 63.7% |
| FCN ResNet-50 | 134 MB | 125 MB | 1.8.0 | 12 | 65.0% |
| FCN ResNet-50-int8 | 34 MB | 29 MB | 1.8.0 | 12 | 64.7% |
- PyTorch Torchvision FCN ResNet50 ==> ONNX FCN ResNet50
- PyTorch Torchvision FCN ResNet101 ==> ONNX FCN ResNet101
- ONNX FCN ResNet50 ==> Quantized ONNX FCN ResNet50
The input is expected to be an image with the shape (N, 3, height, width) where N is the number of images in the batch, and height and width are consistent across all images.
The images must be loaded in RGB with a range of [0, 1] per channel, then normalized per-image using mean = [0.485, 0.456, 0.406] and std = [0.229, 0.224, 0.225].
This model can take images of different sizes as input. However, it is recommended that the images are resized such that the minimum size of either edge is 520.
The following code shows an example of how to preprocess a demo image:
from PIL import Image
from torchvision import transforms
preprocess = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
img = Image.open('dependencies/000000017968.jpg')
img_data = preprocess(img).detach().cpu().numpy()The model has two outputs, ("out", "aux"). "out" is the main classifier and has shape (N, 21, height, width). Each output pixel is one-hot encoded, i.e. np.argmax(out[image, :, x, y]) is that pixel's predicted class. Class 0 is the background class.
"aux" is an auxilliary classifier with the same shape performing the same functionality. The difference between the two is that "out" sources features from last layer of the ResNet backbone, while "aux" sources features from the second-to-last layer.
The following code shows how to overlay the segmentation on the original image:
from PIL import Image
from matplotlib.colors import hsv_to_rgb
import numpy as np
import cv2
classes = [line.rstrip('\n') for line in open('voc_classes.txt')]
num_classes = len(classes)
def get_palette():
# prepare and return palette
palette = [0] * num_classes * 3
for hue in range(num_classes):
if hue == 0: # Background color
colors = (0, 0, 0)
else:
colors = hsv_to_rgb((hue / num_classes, 0.75, 0.75))
for i in range(3):
palette[hue * 3 + i] = int(colors[i] * 255)
return palette
def colorize(labels):
# generate colorized image from output labels and color palette
result_img = Image.fromarray(labels).convert('P', colors=num_classes)
result_img.putpalette(get_palette())
return np.array(result_img.convert('RGB'))
def visualize_output(image, output):
assert(image.shape[0] == output.shape[1] and \
image.shape[1] == output.shape[2]) # Same height and width
assert(output.shape[0] == num_classes)
# get classification labels
raw_labels = np.argmax(output, axis=0).astype(np.uint8)
# comput confidence score
confidence = float(np.max(output, axis=0).mean())
# generate segmented image
result_img = colorize(raw_labels)
# generate blended image
blended_img = cv2.addWeighted(image[:, :, ::-1], 0.5, result_img, 0.5, 0)
result_img = Image.fromarray(result_img)
blended_img = Image.fromarray(blended_img)
return confidence, result_img, blended_img, raw_labels
conf, result_img, blended_img, raw_labels = visualize_output(orig_tensor, one_output)The FCN models have been pretrained on the COCO train2017 dataset, using the subset of classes from Pascal VOC classes. See the Torchvision Model Zoo for more details.
Pretrained weights from the Torchvision Model Zoo were used instead of training these models from scratch. A conversion notebook is provided.
Mean IoU (intersection over union) and global pixelwise accuracy are computed on the COCO val2017 dataset. Torchvision reports these values as follows:
| Model | mean IoU (%) | global pixelwise accuracy (%) |
|---|---|---|
| FCN ResNet 50 | 60.5 | 91.4 |
| FCN ResNet 101 | 63.7 | 91.9 |
If you have the COCO val2017 dataset downloaded, you can confirm updated numbers using the provided notebook:
| Model | mean IoU | global pixelwise accuracy |
|---|---|---|
| FCN ResNet 50 | 65.0 | 99.6 |
| FCN ResNet 50-int8 | 64.7 | 99.5 |
| FCN ResNet 101 | 66.7 | 99.6 |
The more conservative of the two estimates is used in the model files table.
Compared with the fp32 FCN ResNet 50, FCN ResNet 50-int8's mean IoU drop ratio is 0.46% global pixelwise accuracy drop ratio is 0.10% and performance improvement is 1.28x.
Note
The performance depends on the test hardware. Performance data here is collected with Intel® Xeon® Platinum 8280 Processor, 1s 4c per inst ance, CentOS Linux 8.3, data batch size is 1.
FCN ResNet 50-int8 is obtained by quantizing fp32 FCN ResNet 50 model. We use Intel® Neural Compressor with onnxruntime backend to perform quantization. View the instructions to understand how to use Intel® Neural Compressor for quantization.
onnx: 1.9.0 onnxruntime: 1.8.0
wget https://github.com/onnx/models/raw/main/vision/object_detection_segmentation/fcn/model/fcn-resnet50-12.onnxMake sure to specify the appropriate dataset path in the configuration file.
bash run_tuning.sh --input_model=path/to/model \ # model path as *.onnx
--config=fcn_rn50.yaml \
--data_path=path/to/coco/val2017 \
--label_path=path/to/coco/annotations/instances_val2017.json \
--output_model=path/to/save-
Jonathan Long, Evan Shelhamer, Trevor Darrell; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 3431-3440
-
This model is converted from the Torchvision Model Zoo, originally implemented by Francisco Moss here.
- Jack Duvall
- mengniwang95 (Intel)
- airMeng (Intel)
- ftian1 (Intel)
- hshen14 (Intel)
MIT License