TL;DR
This paper presents a novel approach for recognizing real-world objects from light intensity data by integrating edge finding, segmentation, and recognition into an interactive system. By allowing these modules to communicate, the system can efficiently locate objects by prioritizing reliable edges and retrying with less reliable ones if initial recognition fails.
What Problem Does This Paper Address?
Recognizing objects in complex, real-world scenes using only light intensity data is a significant challenge. Unlike structured environments, everyday scenes offer less constraint, making it difficult to infer object identity from individual pixel values. Traditional approaches often treat feature extraction and recognition as separate, sequential steps. However, this paper argues that these modules should interact. For instance, if an initial attempt to recognize an object fails using strong edges, the system should be able to search for weaker edges to aid recognition. This is particularly crucial for curved objects, where edge extraction and interpretation are more complex than for polyhedra.
Methods
The proposed system comprises four main modules:
- Edge Finding: This module extracts edges from light intensity data. It employs two operators: a two-dimensional operator for general edge detection and a one-dimensional operator for faster detection when edge direction is known. These operators detect discontinuities in both light intensity and its gradient. The system classifies edge points into three types (B, L, R) based on the nature of the intensity change. A key innovation is the use of a "reference map" to speed up the search for edge kernels (regions of similar edge points) and a tracking mechanism to extend these kernels into smooth, continuous edges.
- Segmentation of Edges: Once edges are found, they are segmented into straight lines or elliptic curves. This involves two steps: first, dividing an edge into segments based on curvature, and second, approximating each segment with a straight line or an elliptic curve. The curvature is used to identify potential "knots" where the edge direction changes significantly. For approximation, the paper uses Deming's method, which accounts for errors in both x and y coordinates of edge points, to fit lines and curves.
- Recognition of Objects: This module matches the segmented lines and curves to pre-defined object models. The paper mentions two approaches: graph matching (where objects are described by graphs of edges and their relations) and a procedural approach (where specific procedures are defined for recognizing each object). The system can use raw data if necessary for recognition.
- Supervisor: This module orchestrates the interaction between the edge finding, segmentation, and recognition modules. It controls the parameters for edge finding (e.g., reliability level) and decides when to retry recognition with different edge properties. This interactive feedback loop is central to the system's efficiency and robustness.
Datasets
Not discussed in the paper. The paper focuses on demonstrating the system's capabilities with examples of locating a lamp, a book stand, and a telephone in a desk scene.
Key Insights and Ideas
- Interactive System Design: The core insight is that edge finding, segmentation, and recognition should not be isolated processes but should interact. This allows for a more robust and efficient recognition system that can adapt its search strategy based on intermediate results.
- Prioritizing Reliable Edges: The system intelligently prioritizes finding and segmenting the most reliable edges first. If recognition is successful, it saves computation. If not, it can then search for less reliable edges, increasing the chances of successful recognition.
- Handling Curved Objects: The methods for edge finding and segmentation are specifically designed to handle curved objects, which are more challenging than polyhedra. The use of curvature for segmentation and fitting elliptic curves are key to this.
- Reference Map for Efficiency: The use of a reference map to guide the edge kernel search significantly speeds up the edge finding process.
Key Results
The paper demonstrates the system's ability to locate a lamp, a book stand, and a telephone in a desk scene using only light intensity data. The example in Figure 10 shows the original scene, the found edges, the fitted lines, and the lines used for recognition, successfully identifying the three objects. The computing time for locating these three objects in a 256x256 image was approximately 4 minutes, with over half of that time spent on edge finding. The authors note that a system organization that prioritizes finding specific objects is more efficient than one that tries to find all possible objects at once.
Limitations and Future Work
The authors acknowledge that the recognition module (module 3) has less generality at present and is described briefly. Future work includes:
- Zooming-up Capability: Developing a facility for zooming up regions of interest in the scene to perform further investigation when recognition is uncertain.
- Efficient Model Creation: Investigating efficient ways to create models for complex objects.
- Generalizing Recognition: Expanding the generality of the recognition module to handle a wider variety of objects and scenes.
- Handling Doubtful Objects: Further research is needed to check on doubtful objects by zooming up regions of interest.
Summary updated