Critical Response On Prof. Takeo Kanade’s Research Work On Discovering Objects From Images

In this paper, the author Prof. Takeo Kanade and his colleagues have tried to unify the inference or information collected by a person or a robot from multiple visual images of Activities of Daily Living (ADL). The major two objectives of the paper are basically two-fold: firstly, identifying objects instances which are distinct from each other; secondly, extracting the spatial coordinates from the captured images.

Several related research work on discovering objects from images under unsupervised scenario has been done. However, these procedures lack in handling more than one class of objects at a single time and is useful only in cases having lesser variance among the category.

In this research paper, the author has used the bottom up approach of image segmentation for identification and extraction of distinct image. Here appearance consistency and geometric consistency is used as the portion belonging to same objects within two images shows higher level of correlation.

In the algorithm used for ADL analysis, instead of reliability on a single segmentation output, multiple segmentation output is combined together to help in capturing minute details of the objects. In this paper, 25 chunks per image or 4390 chunks in total is used for analysis. The drawback in using a large chunk size is that it also increases the portions unrelated to the objects for which the chunks that are mutually consistent are only used for analysis.

During computation of pairwise consistency, two major factors needs to be taken into consideration: firstly, attributes like ‘texture’, ‘color’, ’shape’ are used in case the appearance differs in different images; secondly, a modified measure of consistency is determined to differentiate between image with high or no texture. Then the mutually consistent portions identified are grouped together. The chunks are then refined to determine the correlation of the pixels with the objects in the image. The extent of co-occurrence is computed and if it exceeds 80%, then a new group of portions is composed. In the analysis, object candidates are computed by considering the cases where the group purity is at least 80%, computed as percentage of correct chunks belonging to the same object. Then the ratio of precision to recall is calculated for all objects. The impact of different parameters impacting this ratio is also analyzed. To evaluate the universal applicability of the program, besides ADL, it was also evaluated on CMU dataset and Flikr images.

CMU dataset has high clutter, disturbances and changes in viewpoint. The study conducted on 10000 segments from 500 images showed that this program out-performs other two models i.e. Russel’s and Kim’s model. In case of Flikr data sets which has inbuilt disturbances from images which are not related to each other, 100 images from flickr.com was analyzed and the result shows that in this case too, the program is superior compared to other models.

Critique

The paper has a very straightforward and plain-spoken approach written in a crisp language which makes it very comprehensible. The presentation of ideas and the analysis given were easy to follow and understand. The theoretical specifics mentioned by the authors are very insightful and help in complementing the elegance of the ideas formulated in the paper.

The manner in which the approach takes into consideration three different data sets: namely, Activities of Daily Living (ADL), CMU dataset and Flikr images indeed makes this study more relevant. As computing this study across multiple data sets makes the output results credible.

Using multiple segments for analysis, consistent images are identified, grouped and further refined for analysis. The approach used by the paper has also been compared with earlier existing baseline systems which enhances the comprehensibility of the research work.

However, being a just 8-page research paper has its own drawbacks too. Clearly, it becomes apparent to the reader that the paper being very crisp in its formulations, lacks the in-depth details, theories and conjectures that could have given the paper more substance.

Ideas for Follow-On Work

The unified approach for identifying objects using three different kind of data sets can undergo further research and empirical and statistical analysis to broaden the field of Image analysis using segmentation technique.

The algorithms proposed, the theorems and the other metrics used in the research paper can be further researched to produce a base for future hypothesis and proposals of discovering objects from images which are related to each other.

This model can be further explored in future for implementing the segmentation analysis for other categories of objects or various other data sets having a wider range of variation in details.