Główne pojęcia
The proposed Clicks2Line algorithm adaptively uses either clicks or lines as user input to efficiently segment elongated objects in interactive image segmentation tasks.
Streszczenie
The paper presents the Clicks2Line algorithm, an interactive image segmentation method that adaptively uses either clicks or lines as user input. The key insights are:
- Existing click-based interactive segmentation methods require a substantial number of clicks to segment elongated regions, as a single click can only cover a limited area.
- Lines can represent elongated regions more effectively than multiple clicks, as a single line can cover a larger area.
- The Clicks2Line algorithm determines whether to use a click or a line as input based on the aspect ratio of the target region. If the aspect ratio is smaller than a threshold, a click is used; otherwise, a line is used.
- The line generation process involves creating multiple line candidates and selecting the optimal line that best represents the target region, considering both the length of the line and the penalty for crossing opposite label regions.
- Experimental results on the GrabCut and Berkeley datasets show that the proposed Clicks2Line algorithm outperforms existing click-based methods, especially in terms of the number of clicks required to achieve high segmentation accuracy for elongated regions.
Statystyki
The paper reports the following key metrics:
Number of Clicks (NoC) at 85%, 90%, and 95% IoU thresholds on the GrabCut and Berkeley datasets.
Comparison of the proposed Clicks2Line algorithm with SimpleClick and MFP methods.
Cytaty
"Although existing click-based methods yield decent segmentation results, they require substantial amount of user clicks to segment long regions as in the case of Figure 1(a)."
"Generally, a line can cover more pixels than a click, thus it provides much better representations of elongated objects. Therefore, one could expect that a single line, as in Figure 1(b), could better indicate a target's appearance than multiple clicks."