FAQs and answers to existing saliency Computing Models)

Ali borji and others published a comprehensive paper on the Significance model on eccv.

Original article from: Ali borji, Dicky N. sihite, and Laurent itti,
"Salient object detection: A Benchmark",

Summarizes the (code or saliency map) that is easily available since and before January 1, 2011, with good accuracy or a model with a high adoption rate.

The full text shares five databases (these five data are all databases containing objects, and the focus of this article is to compare the models in the databases containing objects ).

Human inter-Observer, calculate the inconsistency between this person and other people's labels.

This article answers the following questions:

Question 1: Is there room for improvement in the current model?

A: Since the I/O model is the upper bound of all models, it is usually highly consistent when a significant object is labeled. The existing models and I/O models differ greatly from each other in all databases (the five databases proposed in this article are the same below), so there is still room for improvement.

Question 2: Is there a theoretical lower limit on the accuracy of existing models?

A: Yes. Input a significant value based on the even distribution of all pixels in the image. The AUC calculated from the obtained conspicuous image is 0.5, which is the theoretical lower bound. The AUC value of all models is greater than or equal to this value.

Question 3: What are the main types of models?

A: The current methods are mainly divided into two categories. 1) Prediction Based on the watching point model; 2) extraction and segmentation based on the model of significant objects. On average, the performance of the model based on the gaze point is worse than that based on the object's significance model. The best model in watching point prediction is better than the worst model in object-based model.

Question 4: Why are the two models having different performance on data that is easy to split?

A: The reason is the number of true and false positives. The segmentation algorithm tries to generate a significant white area to include more true positives. On the one hand, the prediction model of gaze points is highly selective and rarely produces false positive points (because the number of gaze points in the figure is relatively small ). In separate experiments, the performance of the prediction model on the monitoring point database is better than that on the segmentation-based model.

Question 5: Is the result of linear combination of existing models better than that of a separate model?

A: combining the best model on each database may lead to better results than all other models. The result of combining two models is similar to that of the three models with the best combination (a little less ).

Question 6: Does the size of an object in an image affect the accuracy of the model?

A: It is indeed challenging to achieve high accuracy on small objects. If 80% of a graph is an object, a model uses the entire graph as a prominent graph to get 80% accuracy. Most models have high accuracy in large-scale object graphs.

Question 7: Does the consistency of manual labeling affect the accuracy of the model?

A: The more consistent the manual annotation, the higher the accuracy of the model.

Question 8: Does each model have similar accuracy for all images in the same database?

A: each model has its own diagram that is the easiest to process and the most difficult to process. In general, there are very vivid objects in the center, and the surrounding images are completely different backgrounds. They are the easiest charts for all models to process. The most difficult thing to deal with is the images with complex textures and objects that contain several different parts, or objects that can attract top-down attention (such as text, face, human body, social behavior, attention orientation, and animals ).

Question 9: Does Gaussian blur on a conspicuous image affect accuracy?

A: Gaussian models have a slight impact on accuracy, but their Qualitative trends and order of models are consistent.

Question 10: Why does the Gaussian Model Change the watching point prediction model without changing the accuracy of the object detection model?

A: There are two possible reasons: 1) there is inconsistency on the gaze point, and the model results usually fall into the attention point. Therefore, the Gaussian model improves this result; 2) in the detection of notable objects, the evaluation index is calculated in an image region, while in the prediction model of the gaze point, it is usually calculated on the sample graph of the Eye watching point. For the former, Gaussian Blur is only at the left and right of the edge, while for the latter, the model results can also be improved.

Question 11: Does adding a central Offset provide accuracy of the model?

A: All databases have the central offset attribute. Adding a center offset can improve the model with poor performance, but it has the opposite effect on the model with better performance.

Question 12: Are there any similarity between the results of the existing model?

A: using linear correlation coefficient, we can conclude that the existing model does have similarity before.

Question 13: What is the relationship between consistency between models and that between manual labeling?

A: The relationships between them can be summarized as follows: 1) for images with both model consistency and manual labeling consistency, they usually contain clear objects, the background color is different from the color of the object. 2) for images with inconsistent human tags and different model results, most of them are those that are not easy to define significant objects. These images have complex texture backgrounds, and objects contain several different parts. 3) images with different models are usually complex in the background texture, and the features of notable objects and backgrounds are similar. 4) Finally, for those images with the same model but different models, the objects in the graph usually contain multiple parts, and each part has different features from the background. In general, there are few images with manual inconsistencies.

FAQs and answers to existing saliency Computing Models)