A perceptually motivated online benchmark for image matting

Cosegmentation Revisited: Models and Optimization

“…In section 4.1, we compare the performance of the different optimization methods, and in section 4.2, using the best optimization procedure for each model, we compare the performance and robustness of such models. Dataset Given the difficulty in acquiring ground truth data for the cosegmentation problem, we used composites of 40 different backgrounds with 20 foreground objects from the database in [20], for which high quality alpha mattes are available. The database in [20] has more than 20 images; we selected objects with fewer transparencies.…”

“…Dataset Given the difficulty in acquiring ground truth data for the cosegmentation problem, we used composites of 40 different backgrounds with 20 foreground objects from the database in [20], for which high quality alpha mattes are available. The database in [20] has more than 20 images; we selected objects with fewer transparencies. Representative images out of these 20 pairs are shown in Fig.…”

Confidence-driven image co-matting

“…The recent work by Chen et al [13] integrates sampling-based priors with both local and nonlocal affinities together to achieve the best performance on the benchmark dataset [14]. However, when dealing with real-world image with highly complex appearance, both sampling priors and affinity terms are error-prone, and combining them together does not help much.…”

“…In our case, for training, we randomly collect 100 background images from internet and blend them with the foreground images using their corresponding ground-truth alpha mattes provided by the matting evaluation benchmark [14]. Global Sampling matting is then applied on all the synthesized images to estimate alpha mattes, and features are then extracted as articulated above.…”

“…We then perform model evaluation 5 times, by selecting one of the dividends for testing and all others for training each time. It takes around 4.5 hours to extract˜10K features from all 1700 images(100 background image blended with 17 foreground ones drawn from the matting evaluation benchmark [14]). Once the features are generated, the process of cross-validation is quite fast, taking about 3.5 minutes.…”