angtian / nemo Goto Github PK

BBox2D.include in BboxTools 1.0.2 seems to only accept single points or a Bbox2D object (code pasted below). points_2d is a numpy array of points, which causes an error at the line above, so generate3Dpascal3D.py is failing to generate any annotations.

def include(self, other):
        """
        Check if other is inside this box. Notice include means strictly include, other could not place at the boundary
        of this bbox.
        :param other: (Bbox2D or tuple of int) bbox or point
        :return: (bool) True or False
        """
        if type(other) == Bbox2D:
            out = True
            for i in range(2):
                if self.bbox[i][0] > other.bbox[i][0]:
                    out = False
                if self.bbox[i][1] < other.bbox[i][1]:
                    out = False
            return out

        if type(other) == tuple and len(other) == 2:
            if other[0] < self.bbox[0][0] or other[0] >= self.bbox[0][1]:
                return False
            if other[1] < self.bbox[1][0] or other[1] >= self.bbox[1][1]:
                return False
            return True
        raise Exception('Include method suppose to be point or bbox, but got %s' % str(other))

Hi,

Thanks for the great work and sharing the code!

I've got one question about the non-occluded split of Pascal3D+. In your paper, StarMap gets 89% Acc30 while the author of StarMap only reported an accuracy at 82% in their paper, I suppose this difference is resulted from a different dataset split for evaluation. Could you tell me how you get the non-occluded test images on Pascal3D+ and is there any difference between the images in occlusion L0, L1, L2, and L3?

Hi Angtian,

As the title suggests, I was just wondering whether you think it's possible and whether you have tried to train NeMo on KITTI3D considering that the samples contain ground truth only for azimuth and not for elevation or theta.

Hello,

In the paper, while training the Neural Mesh Models, "the correspondence between the feature vector f_i in the feature map F and the vector θ_r on the neural mesh model is given by the 2D projection of the mesh with camera parameters m". I assume that this means during training, the neural mesh model is rendered using the ground-truth camera pose m, then compared with the features of the RGB image.

However, I don't see this being the case in the code. Can you give some comments about the actual implementation?

Thanks in advance.