scenedino/downstream_head/crf.py

#
# Authors: Wouter Van Gansbeke & Simon Vandenhende
# Licensed under the CC BY-NC 4.0 license (https://creativecommons.org/licenses/by-nc/4.0/)

import numpy as np
import pydensecrf.densecrf as dcrf
import pydensecrf.utils as utils
import torch
import torch.nn.functional as F
import torchvision.transforms.functional as VF

MAX_ITER = 10
POS_W = 3
POS_XY_STD = 0.3  # 1
Bi_W = 4
Bi_XY_STD = 20  # 67
Bi_RGB_STD = 3


def dense_crf(image_tensor: torch.FloatTensor, output_logits: torch.FloatTensor):
    image = np.array(VF.to_pil_image(image_tensor))[:, :, ::-1]
    H, W = image.shape[:2]
    image = np.ascontiguousarray(image)

    output_logits = F.interpolate(output_logits.unsqueeze(0), size=(H, W), mode="bilinear",
                                  align_corners=False).squeeze()
    output_probs = F.softmax(output_logits, dim=0).cpu().numpy()

    c = output_probs.shape[0]
    h = output_probs.shape[1]
    w = output_probs.shape[2]

    U = utils.unary_from_softmax(output_probs)
    U = np.ascontiguousarray(U)

    d = dcrf.DenseCRF2D(w, h, c)
    d.setUnaryEnergy(U)
    d.addPairwiseGaussian(sxy=POS_XY_STD, compat=POS_W)
    d.addPairwiseBilateral(sxy=Bi_XY_STD, srgb=Bi_RGB_STD, rgbim=image, compat=Bi_W)

    Q = d.inference(MAX_ITER)
    Q = np.array(Q).reshape((c, h, w))
    return Q