Crawl approach for creation and automatic annotation of custom datasets
A method of creating custom datasets is provided. The method comprises capturing, with a camera, a sequence of images of a surface in a specified crawl direction over a number of sequential crawl steps that represent changes of camera position. For each crawl step, the method matches a first image and a second image captured at sequential camera positions. The method models each crawl step as a pairwise quadratic transformation between the first image and second image according to a number of respective anchor points in each image selected as best candidates to fit a regression model.
1 . A computer-implemented method of creating custom datasets, the method comprising:
using a number of processors to perform the operations of:
capturing, with a camera, a sequence of images of a surface in a specified crawl direction over a number of sequential crawl steps that represent changes of camera position;
for each crawl step, matching a first image and a second image captured at sequential camera positions; and
modeling each crawl step as a pairwise quadratic transformation between the first image and second image according to a number of respective anchor points in each image selected as best candidates to fit a regression model.
2 . The method of claim 1 , wherein matching the first image and second image further comprises:
dividing each image into a number of tiles;
matching the tiles; and
generating the anchor points.
3 . The method of claim 1 , wherein the image captured at each crawl step is captured under a number of different lighting conditions.
4 . The method of claim 3 , wherein the anchor points are selected according to a statistical filter for a distribution of position of the anchor points across the different lighting conditions.
5 . The method of claim 3 , wherein the anchor points are selected according to distribution of direction of movement of the anchor points across the different lighting conditions.
6 . The method of claim 1 , further comprising:
receiving input of a bounding box around an object on the surface; and
predicting a trajectory of the vertices of the bounding box over the sequence of images according to the crawl step modeling.
7 . The method of claim 6 , further comprising using a unique identifier of the object to filter any duplicate images.
8 . The method of claim 6 , wherein specifying the trajectory of the vertices of the bounding box automatically annotates the sequence of images in all lighting conditions, and wherein truth data only needs to be provided for one image in the sequence.
9 . The method of claim 6 , further comprising:
responsive to a determination that the object is located at the start of the sequence of images, performing forward propagation;
responsive to a determination that the object is located at the end of the sequence of images, performing backward propagation; and
responsive to a determination that the object is located at a position other than the start or end of the sequence of images, performing both forward and backward propagation.
10 . A system for creating custom datasets, the system comprising:
a storage device that stores program instructions;
one or more processors operably connected to the storage device and configured to execute the program instructions to cause the system to:
capture, with a camera, a sequence of images of a surface in a specified crawl direction over a number of sequential crawl steps that represent changes of camera position;
for each crawl step, match a first image and a second image captured at sequential camera positions; and
model each crawl step as a pairwise quadratic transformation between the first image and second image according to a number of respective anchor points in each image selected as best candidates to fit a regression model.
11 . The system of claim 10 , wherein matching the first image and second image further comprises:
dividing each image into a number of tiles;
matching the tiles; and
generating the anchor points.
12 . The system of claim 10 , wherein the image captured at each crawl step is captured under a number of different lighting conditions.
13 . The system of claim 12 , wherein the anchor points are selected according to a statistical filter for a distribution of position of the anchor points across the different lighting conditions.
14 . The system of claim 12 , wherein the anchor points are selected according to distribution of direction of movement of the anchor points across the different lighting conditions.
15 . The system of claim 10 , wherein the processors further execute program instructions to cause the system to:
receive input of a bounding box around an object on the surface; and
predict a trajectory of the vertices of the bounding box over the sequence of images according to the crawl step modeling.
16 . The system of claim 15 , wherein the processors further execute program instructions to cause the system to use a unique identifier of the object to filter any duplicate images.
17 . The system of claim 15 , wherein specifying the trajectory of the vertices of the bounding box automatically annotates the sequence of images in all lighting conditions, and wherein truth data only needs to be provided for one image in the sequence.
18 . The system of claim 15 , wherein the processors further execute program instructions to cause the system to:
responsive to a determination that the object is located at the start of the sequence of images, perform forward propagation;
responsive to a determination that the object is located at the end of the sequence of images, perform backward propagation; and
responsive to a determination that the object is located at a position other than the start or end of the sequence of images, perform both forward and backward propagation.
19 . A computer program product for creating custom datasets, the computer program product comprising:
a computer-readable storage medium having program instructions embodied thereon to perform the operations of:
capturing, with a camera, a sequence of images of a surface in a specified crawl direction over a number of sequential crawl steps that represent changes of camera position;
for each crawl step, matching a first image and a second image captured at sequential camera positions; and
modeling each crawl step as a pairwise quadratic transformation between the first image and second image according to a number of respective anchor points in each image selected as best candidates to fit a regression model.
20 . The computer program product of claim 19 , the computer-readable storage medium further comprising instructions for:
receiving input of a bounding box around an object on the surface; and
predicting a trajectory of the vertices of the bounding box over the sequence of images according to the crawl step modeling.