Landmark Localization via Visual Search Voting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional sliding window approaches for object detection in images, such as face detection, struggle with variations like profile views, slanted views, occluded views, and detecting landmarks, due to difficulties in learning a good classifier encoding all possible variations.
Innovation Solution
The method involves estimating landmark locations in a test image by using a similarity voting approach, where known landmark locations in object images vote for corresponding locations in the test image, generating a voting map to determine the most likely positions of landmarks like eyes, nose, mouth, and chin, and averaging these estimates for accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a sliding window approach with binary classifier is used for object detection, then the detection process can be implemented with a systematic search method, but the detection accuracy deteriorates for profile views, slanted views, and occluded views of faces
Solution Approach 1:
The patent creates a detection model by copying and adapting features from training images that contain various face orientations and occlusions. The trained classifier learns to recognize patterns across different views by copying successful detection strategies from the training data, enabling accurate detection of profile, slanted, and occluded faces without requiring separate detection mechanisms for each view type.
2Adaptability or versatility
If a binary classifier is trained to encode all possible variations of face views, then the classifier can potentially detect various face orientations, but the complexity of training and computing resources required increases significantly
Solution Approach 1:
The patent segments the face detection problem into multiple stages: first detecting the presence of a face using a trained classifier, then refining the detection by analyzing local features and gradients within the detected region. This segmentation allows the system to handle various face variations without requiring the entire classifier to encode every possible detail of each view type, reducing overall complexity while maintaining versatility.
Solution Approach 2:
The patent performs preliminary actions during the training phase by pre-processing training images to include various face orientations, occlusions, and lighting conditions. The classifier is pre-trained on this diverse dataset to learn robust features that generalize across different views. This preliminary action during training reduces the need for complex real-time adjustments during actual detection, simplifying the operational complexity while maintaining high adaptability.
3Adaptability or versatility
If the sliding window approach is used with varying scales, then the detection can adapt to different face sizes in the image, but the computational time and processing overhead increase
Solution Approach 1:
The patent implements dynamic scaling by adjusting the window size and stride based on the detected face size and image resolution. Rather than uniformly testing all possible scales with fixed step sizes, the system dynamically adapts the search parameters to the specific characteristics of each image and detected region, reducing unnecessary computations while maintaining the ability to detect faces at various scales.
Data Source
AI summary
One exemplary embodiment involves identifying feature matches between each of a plurality of object images and a test image, each of the feature matches between a feature of a respective object image and a matching feature of the test image, wherein there is a spatial relationship between each respective object image feature and a first landmark of the object image, the first landmark at a known location in the object image. The embodiment additionally involves estimating a plurality of locations for a second landmark for the test image, the estimated locations based at least in part on the feature matches and the spatial relationships, and estimating a final location for the second landmark from the plurality of locations for the second landmark for the test image.