external/WildCamera/tools/calibrator.py

import numpy as np
import torch
from einops import rearrange
from torchvision import transforms
from tools.tools import resample_rgb
from tools.visualization import tensor2rgb

class MonocularCalibrator(torch.nn.Module):
    def __init__(self, l1_th=0.02):
        """ Calibrate Camera Intrinsic from Incidence Field.

        Args:
            l1_th (float): RANSAC Inlier Count Threshold. Default 0.02.
            RANSAC_num (int): RANSAC Random Sampling Point Number. Default: 20000.
        """

        super().__init__()
        self.RANSAC_num = 20000
        self.l1_th = l1_th

    def initcoords2D(self, b, h, w, device, homogeneous=False):
        """ Init Normalized Pixel Coordinate System
        """

        query_coords = torch.meshgrid(
            (
                torch.linspace(-1 + 1 / h, 1 - 1 / h, h, device=device),
                torch.linspace(-1 + 1 / w, 1 - 1 / w, w, device=device),
            ),
            indexing='ij'
        )
        query_coordsx, query_coordsy = query_coords[1], query_coords[0]

        if homogeneous:
            query_coords = torch.stack((query_coordsx, query_coordsy, torch.ones_like(query_coordsx)), dim=0).view([1, 3, h, w]).expand([b, 3, h, w])
        else:
            query_coords = torch.stack((query_coordsx, query_coordsy), dim=0).view([1, 2, h, w]).expand([b, 2, h, w])
        return query_coords

    @staticmethod
    def norm_intrinsic(intrinsic, b, h, w, device):
        """ Map Intrinsic to Normalized Image Coordinate System [-1, 1]
        """
        scaleM = torch.eye(3).view([1, 3, 3]).expand([b, 3, 3]).to(device)
        scaleM[:, 0, 0] = float(1 / w) * 2
        scaleM[:, 1, 1] = float(1 / h) * 2

        scaleM[:, 0, 2] = -1.0
        scaleM[:, 1, 2] = -1.0
        return scaleM @ intrinsic

    @staticmethod
    def unnorm_intrinsic(intrinsic, b, h, w, device):
        """ Unmap Intrinsic to Image Coordinate System [0.5, h / w - 0.5]
        """
        scaleM = torch.eye(3).view([1, 3, 3]).expand([b, 3, 3]).to(device)
        scaleM[:, 0, 0] = float(1 / w) * 2
        scaleM[:, 1, 1] = float(1 / h) * 2

        scaleM[:, 0, 2] = -1.0
        scaleM[:, 1, 2] = -1.0
        return scaleM.inverse() @ intrinsic

    def intrinsic2incidence(self, intrinsic, b, h, w, device):
        """ Compute Gt Incidence Field from Intrinsic
        """
        coords3d = self.initcoords2D(b, h, w, device, homogeneous=True)
        intrinsic = MonocularCalibrator.norm_intrinsic(intrinsic, b, h, w, device)

        intrinsic = intrinsic.view([b, 1, 1, 3, 3])
        coords3d = rearrange(coords3d, 'b d h w -> b h w d 1')
        coords3d = torch.linalg.inv(intrinsic) @ coords3d
        coords3d = rearrange(coords3d.squeeze(-1), 'b h w d -> b d h w')
        normalray = torch.nn.functional.normalize(coords3d, dim=1)
        return normalray

    def scoring_function_xy(self, normal_RANSAC, normal_ref):
        """ RANSAC Scoring Function
        """
        xx, yy, _ = torch.split(normal_RANSAC, 1, dim=1)
        xxref, yyref, zzref = torch.split(normal_ref, 1, dim=0)
        xxref = xxref / zzref
        yyref = yyref / zzref

        diffx = torch.sum((xx - xxref.unsqueeze(0)).abs() < self.l1_th, dim=[1, 2])
        diffy = torch.sum((yy - yyref.unsqueeze(0)).abs() < self.l1_th, dim=[1, 2])

        return diffx, diffy

    def get_sample_idx(self, h, w, prob=None, seed=None):
        if seed is not None:
            np.random.seed(seed)

        if prob is not None:
            prob = prob.view([1, int(h * w)]).squeeze().cpu().numpy()
            sampled_index = np.random.choice(
                np.arange(int(h * w)),
                size=self.RANSAC_num,
                replace=False,
                p=prob,
            )
        else:
            sampled_index = np.random.choice(
                np.arange(int(h * w)),
                size=self.RANSAC_num,
                replace=False,
            )
        return sampled_index

    def sample_wo_neighbour(self, x, sampled_index):
        assert len(x) == 1
        _, ch, h, w = x.shape
        x = x.contiguous().view([ch, int(h * w)])
        return x[:, sampled_index]

    def minimal_solver(self, coords2Ds, normalrays, RANSAC_trial):
        """ RANSAC Minimal Solver
        """
        minimal_sample = 2
        device = coords2Ds.device

        sample_num = int(minimal_sample * RANSAC_trial)
        coords2Dc, normal = coords2Ds[:, 0:sample_num], normalrays[:, 0:sample_num]

        x1, y1, _ = torch.split(coords2Dc, 1, dim=0)
        n1, n2, n3 = torch.split(normal, 1, dim=0)

        n1 = n1 / n3
        n2 = n2 / n3

        x1, y1 = x1.view(minimal_sample, RANSAC_trial), y1.view(minimal_sample, RANSAC_trial)
        n1, n2 = n1.view(minimal_sample, RANSAC_trial), n2.view(minimal_sample, RANSAC_trial)

        fx = (x1[1] - x1[0]) / (n1[1] - n1[0] + 1e-10)
        bx = (x1[0] - n1[0] * fx) * 0.5 + (x1[1] - n1[1] * fx) * 0.5

        fy = (y1[1] - y1[0]) / (n2[1] - n2[0] + 1e-10)
        by = (y1[0] - n2[0] * fy) * 0.5 + (y1[1] - n2[1] * fy) * 0.5

        intrinsic = torch.eye(3).view([1, 3, 3]).repeat([len(fx), 1, 1]).to(device)
        intrinsic[:, 0, 0] = fx
        intrinsic[:, 1, 1] = fy
        intrinsic[:, 0, 2] = bx
        intrinsic[:, 1, 2] = by

        return intrinsic

    def calibrate_camera_4DoF(self, incidence, RANSAC_trial=2048):
        """ 4DoF RANSAC Camera Calibration

        Args:
            incidence (tensor): Incidence Field
            RANSAC_trial (int): RANSAC Iteration Number. Default: 2048.
        """
        # Calibrate assume a simple pinhole camera model
        b, _, h, w = incidence.shape
        device = incidence.device
        coords2D = self.initcoords2D(b, h, w, device, homogeneous=True)

        sampled_index = self.get_sample_idx(h, w)
        normalrays = self.sample_wo_neighbour(incidence, sampled_index)
        coords2Ds = self.sample_wo_neighbour(coords2D, sampled_index)

        # Prepare for RANSAC
        intrinsic = self.minimal_solver(coords2Ds, normalrays, RANSAC_trial)

        valid = (intrinsic[:, 0, 0] < 1e-2).float() + (intrinsic[:, 1, 1] < 1e-2).float()
        valid = valid == 0
        intrinsic = intrinsic[valid]

        # RANSAC Loop
        intrinsic_inv = torch.linalg.inv(intrinsic)
        normalray_ransac = intrinsic_inv @ coords2Ds.unsqueeze(0)
        diffx, diffy = self.scoring_function_xy(normalray_ransac, normalrays)
        intrinsic_x, intrinsic_y = intrinsic, intrinsic

        maxid = torch.argmax(diffx)
        fx, bx = intrinsic_x[maxid, 0, 0], intrinsic_x[maxid, 0, 2]
        maxid = torch.argmax(diffy)
        fy, by = intrinsic_y[maxid, 1, 1], intrinsic_y[maxid, 1, 2]

        intrinsic_opt = torch.eye(3).to(device)
        intrinsic_opt[0, 0] = fx
        intrinsic_opt[0, 2] = bx
        intrinsic_opt[1, 1] = fy
        intrinsic_opt[1, 2] = by

        intrinsic_opt = MonocularCalibrator.unnorm_intrinsic(intrinsic_opt.unsqueeze(0), b, h, w, device)
        return intrinsic_opt.squeeze(0)

    def calibrate_camera_1DoF(self, incidence, r, RANSAC_trial=2048):
        """ 1DoF RANSAC Camera Calibration

        Args:
            incidence (tensor): Incidence Field.
            r: Aspect Ratio Restoration from Network Inference Resolution (480 x 640) to the Original Resolution
            RANSAC_trial (int): RANSAC Iteration Number. Default: 2048.
        """

        # Calibrate assume a simple pinhole camera model
        b, _, h, w = incidence.shape
        assert b == 1
        # r = (scaleM[0, 1, 1] / scaleM[0, 0, 0]).item()
        device = incidence.device
        coords2D = self.initcoords2D(b, h, w, device, homogeneous=True)

        sampled_index = self.get_sample_idx(h, w)
        normalrays = self.sample_wo_neighbour(incidence, sampled_index)
        coords2Ds = self.sample_wo_neighbour(coords2D, sampled_index)

        # Prepare for RANSAC
        fs = torch.linspace(100, 4096, steps=RANSAC_trial)
        intrinsic = torch.eye(3).view([1, 3, 3]).expand([2048, 3, 3]).contiguous().to(device)
        intrinsic[:, 0, 2] = float(w / 2)
        intrinsic[:, 1, 2] = float(h / 2)
        intrinsic[:, 0, 0] = fs
        intrinsic[:, 1, 1] = fs * r
        intrinsic = self.norm_intrinsic(intrinsic, b, h, w, device)

        # RANSAC Loop
        intrinsic_inv = torch.linalg.inv(intrinsic)
        normalray_ransac = intrinsic_inv @ coords2Ds.unsqueeze(0)
        diffx, diffy = self.scoring_function_xy(normalray_ransac, normalrays)

        maxid = torch.argmax(diffx + diffy)
        fx, bx = intrinsic[maxid, 0, 0], intrinsic[maxid, 0, 2]
        fy, by = intrinsic[maxid, 1, 1], intrinsic[maxid, 1, 2]

        intrinsic_opt = torch.eye(3).to(device)
        intrinsic_opt[0, 0] = fx
        intrinsic_opt[0, 2] = bx
        intrinsic_opt[1, 1] = fy
        intrinsic_opt[1, 2] = by

        intrinsic_opt = MonocularCalibrator.unnorm_intrinsic(intrinsic_opt.unsqueeze(0), b, h, w, device)
        return intrinsic_opt.squeeze(0)

    def restore_image(self, image, intrinsic, fixcrop=True):
        # Adjust Intrinsic with Crop and Resize
        w, h = image.size
        wt, ht = image.size

        # Fix Aspect Ratio, Avoid Image Reduced
        resizeM = np.eye(3)
        if intrinsic[0, 0] > intrinsic[1, 1]:
            r = intrinsic[0, 0] / intrinsic[1, 1]
            resizeM[1, 1] = r
            wt, ht = wt, ht * r
        else:
            r = intrinsic[1, 1] / intrinsic[0, 0]
            resizeM[0, 0] = r
            wt, ht = wt * r, ht
        wt, ht = int(np.ceil(wt).item()), int(np.ceil(ht).item())

        # Fix Crop
        cropM = np.eye(3)
        if fixcrop:
            intrinsic_ = resizeM @ intrinsic
            padding_lr, padding_ud = intrinsic_[0, 2] - wt / 2, intrinsic_[1, 2] - ht / 2
            if padding_lr < 0:
                cropM[0, 2] = -padding_lr
            if padding_ud < 0:
                cropM[1, 2] = -padding_ud

            wt, ht = int(np.ceil(wt + np.abs(padding_lr)).item()), int(np.ceil(ht + np.abs(padding_ud)).item())

        resample_matrix = np.linalg.inv(cropM @ resizeM)
        totensor = transforms.ToTensor()
        image_restore = resample_rgb(
            totensor(image).unsqueeze(0),
            torch.from_numpy(resample_matrix).float().view([1, 3, 3]),
            batch=1, ht=ht, wd=wt, device=torch.device("cpu")
        )

        return tensor2rgb(image_restore, viewind=0)