Pose Estimation without tracking, poor perfomance

[dhanushTDS]

When doing pose estimation only and no tracking, what is the suggested format for RGB, Depth, and Mask Images? Currently, I have scraped some of your code together like this, but I am getting poor results.

Also, to clarify what does the pose correspond to : the transformation of object wrt to camera or the other way around

def get_color(image_input):
    color = imageio.imread(image_input)[...,:3]
    color = cv2.resize(color, (480, 640), interpolation=cv2.INTER_NEAREST)
    return color

def get_depth(depth_input):
    depth = cv2.imread(depth_input,-1)/1e3
    depth = cv2.resize(depth, (480, 640), interpolation=cv2.INTER_NEAREST)
    depth[(depth<0.1) | (depth>=np.inf)] = 0
    return depth

def get_mask(mask_input):
    mask = cv2.imread(mask_input,-1)
    if len(mask.shape)==3:
      for c in range(3):
        if mask[...,c].sum()>0:
          mask = mask[...,c]
          break
    mask = cv2.resize(mask, (480, 640), interpolation=cv2.INTER_NEAREST).astype(bool).astype(np.uint8)
    return mask


peg_rgb_image = get_color(rgb_image_path)
depth_image = get_depth(depth_image_path)
mask_image = get_mask(mask_image_path).astype(bool)

cv2.imshow('1', peg_rgb_image)
cv2.waitKey(0)  # Wait indefinitely until a key is pressed
cv2.destroyAllWindows()

pose = est.register(K=rros_camera_I_matrix, rgb=peg_rgb_image, depth=depth_image, ob_mask=mask_image, iteration=est_refine_iter)

center_pose = pose @ np.linalg.inv(to_origin)
vis = draw_posed_3d_box(rros_camera_I_matrix, img=peg_rgb_image, ob_in_cam=center_pose, bbox=bbox)
vis = draw_xyz_axis(peg_rgb_image, ob_in_cam=center_pose, scale=0.1, K=rros_camera_I_matrix, thickness=3, transparency=0, is_input_rgb=True)
# cv2.imshow('1', vis[...,::-1])
cv2.imshow('1', vis)
cv2.waitKey(0)  # Wait indefinitely until a key is pressed
cv2.destroyAllWindows()

[dhanushTDS]

@wenbowen123

[wenbowen123]

Hi, there are typically issues about setting up your data, in particular the depth format. I'd suggest to search in the issues as there are couple related.

The estimated poses are object to the camera.

[utsavrai]

Hi @dhanushTDS and @wenbowen123, in order to perform pose estimation of multiple images, do we need to provide mask for each input rgb image?

[dhanushTDS]

If we do what @ wenbowen123 mentioned, where we are not tracking; instead, we treat each image as a sample from an unordered sequence, then we have to supply the mask to the estimator for each image.

In their example code, they are tracking on frames of a continuous video and only need the mask for the first image.

[utsavrai]

Thanks @dhanushTDS

[wenbowen123]

We just updated the readme, where there is a trouble shooting section. Please check it out.