Semantic Image Matting

Abstract

If it is the author's pre-published version, changes introduced as a result of publishing processes such as copy-editing and formatting may not be reflected in this document. For a definitive version of this work, please refer to the published version.

Boosting Robustness of Image Matting with Context Assembling and Strong Data Augmentation

“…Due to the nature of relying on low-level color cues, their assumptions are easily violated in complex images. To overcome this dilemma, deep matting methods [1,4,11,14,18,19,22,28,29,34,36] appeared with the development of deep learning.…”

“…To verify the generalization ability of matting methods to unseen domains, comparison experiments are carried out in Table 5. Specifically, we test models trained merely with the DIM dataset [34] on several different benchmarks without fitting on their training set (except that SIM [28] is trained on SIMD [28], which has 763 foregrounds, 332 more foregrounds than DIM). The test benchmarks include Distinction-646 [24], SIMD [28], and AIM-500 [17].…”

“…Specifically, we test models trained merely with the DIM dataset [34] on several different benchmarks without fitting on their training set (except that SIM [28] is trained on SIMD [28], which has 763 foregrounds, 332 more foregrounds than DIM). The test benchmarks include Distinction-646 [24], SIMD [28], and AIM-500 [17]. Distinction-646 and SIMD are synthetic benchmarks, and AIM-500 is a real-world one but with simple scenes, so none of them alone is perfect for measuring generalization ability of a model.…”

“…• More visual results on real-world images and benchmarks, including Composition-1k [34], Distinction-646 [24], SIMD our [28], AIM-500 [17].…”

“…Meanwhile, it has been known that deep matting models trained on synthetic data undertake the risk of poor generation to real-world domains [14,28,37] (Fig. 1).…”