CCTV Revisited: Face Analysis Under Challenging Conditions
Motivation: It is difficult to obtain meaningful data from low-resolution images, especially given challenging conditions such as motion blur, sensor noise, pose and illumination changes. Although some studies provide questionable routes to analysing this sort of poor input, thus far very little research has been conducted on such uncontrolled data.
Hypothesis: Landmarks (reference points) and edges provide a strong cue for estimating 3D face shape from 2D images.
Novelty: We adapt the ICP algorithm for use in fitting a 3DMM to image edges automatically. This is the first approach that uses hard model/edge correspondences and leads to an algorithm that is both efficient and robust.
Edges are an attractive feature to exploit because they are relatively insensitive to changes in illumination and camera parameters. They also convey shape and pose information in a rather direct manner.
Our approach is as follows:
- 1. Compute a binary edge map from the input image.
- 2. Given an initial estimate of the shape and pose of the face, compute which vertices lie on the occluding boundary of the estimated shape.
- 3. For each projected edge vertex, find the closest image edge pixel. This can be done efficiently by storing the image edge pixels in a Kd-tree.
- 4. Filter the correspondences by removing matches where the distance to the closest edge pixel is larger than a threshold.
- 5. The edge correspondences provide new 2D landmarks. Return to step 2 with the new shape and pose estimates.
We propose an approach inspired by the iterated closest point (ICP) algorithm, based on computing hard correspondences between model vertices and edge pixels. Three example scans from Basel Face Model fitted using our method are shown below.
