Add eccentricity calculations #57
Conversation
|
[Pre-review comment] It is unlikely that we will merge this branch, due to some hard-coded design decisions (e.g. the aforementioned manual selection of erosion/dilation kernels). When I review this, I will provide 2 types of feedback:
|
| dilated_mask = cv2.dilate(eroded_mask, kernel, iterations=1) | ||
| contours, _ = cv2.findContours(dilated_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE) | ||
|
|
||
| # if len(contours) > 0: |
There was a problem hiding this comment.
Please remove the commented-out code if it's not going to be part of the final version.
| @@ -119,3 +126,57 @@ def calculate_moments(contour_list): | |||
| render = np.zeros([frame_size, frame_size, 1], dtype=np.uint8) | |||
| _ = cv2.drawContours(render, contour_list, -1, [1], -1) | |||
| return cv2.moments(render) | |||
|
|
|||
| def calculate_contours(self, contour_list): | |||
There was a problem hiding this comment.
Function name/docstring could use some additional specificity.
calculate_contoursdoesn't quite describe what this function does. The function appears to do the following:- Render contours provided
- Erode and dilate render
- Calculate new "filtered" contours
- Docstring mismatches almost all fields
- Summary incorrect
- parameters don't include
contour_listand include 3 other values - return listed as
none, but function has return
| mask = np.zeros(frame_size, dtype=np.uint8) | ||
| cv2.drawContours(mask, contour_list, -1, (1), thickness=cv2.FILLED) | ||
| kernel = np.ones((7, 7), np.uint8) | ||
| eroded_mask = cv2.erode(mask, kernel, iterations=1) |
There was a problem hiding this comment.
[Future work]
Erosion/dilation parameters should likely be somehow attached to pixel information.
Here, Lucia uses a 7x7 kernel with 1 iteration with the intent to remove the tail for her dataset (OFA, similar to https://doi.org/10.7910/DVN/SAPNJG). If we want to generalize this, we should probably get an average thickness of tail (in cm) and determine the kernel size via the number of pixel radius to adequately erode a feature of that size.
| :param color: color of segmentation contours rendered on the GUI. | ||
| :return: None | ||
| """ | ||
| frame_size = [800, 800] |
There was a problem hiding this comment.
[Future work]
Frame size is arbitrary and while 800x800 will work for this data, it should probably depend on max of contour data rather than hard-coded values. Would need to think a bit more, but may need a buffer edge with morphological filters (although erode should only shrink, should be fine...).
| @@ -25,6 +25,8 @@ def __init__(self, poses: PoseEstimation, identity: int, | |||
| self._moments = np.zeros((self._poses.num_frames, len(self._moment_keys)), dtype=np.float32) | |||
| self._seg_data = self._poses.get_segmentation_data(identity) | |||
| self._seg_flags = self._poses.get_segmentation_flags(identity) | |||
| self._contours = [None] * self._poses._num_frames | |||
There was a problem hiding this comment.
Name of attribute should better describe the content. These are morphologically filtered contours, as opposed to "contours" (already present in self._seg_data)
| @@ -62,6 +66,9 @@ def get_moment(self, frame, key): | |||
| """ | |||
| key_idx = self._moment_keys.index(key) | |||
| return self._moments[frame, key_idx] | |||
|
|
|||
| def get_contours(self, frame): | |||
| @@ -86,6 +89,17 @@ def per_frame(self, identity: int) -> np.ndarray: | |||
| euler_number[frame] = np.sum(contour_flags[:len(contour_list)]==1)-np.sum(contour_flags[:len(contour_list)]==0) | |||
| hole_area_ratio[frame] = (hole_areas * self._pixel_scale**2)/self._moment_cache.get_moment(frame, 'm00') | |||
|
|
|||
| # CALCULATE ECCENTRICITY | |||
| contours = self._moment_cache.get_contours(frame) | |||
| if len(contours) > 0: | |||
There was a problem hiding this comment.
How does this handle missing segmentation (e.g. the None initialization here)?
You may also need to check if contours is not None and ....
Note: I manually chose a kernel for erosion/dilation, which probably isn't the smartest but it worked for the videos I think