Learning Affinity-Aware Upsampling for Deep Image Matting

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

“…The convolution operations or dynamic kernels learned from local regions [4,22] model local context into the network. On another side, dynamic networks were introduced to matting [4,18,22] to enlarge the model capacity. They also benefit the network in combination with the context assembling [22].…”

“…As the bridge to recover resolutions and capture details, the decoder matters for matting. For instance, previous methods applied feature skip [22], attention-guided refinement [18] or dynamic upsampling [4] to build functional matting decoders aiming at richer details. As the first matting method applying transformers, and considering the importance of the decoder, we investigate an efficient decoder design for our framework.…”

“…Low-Level Feature Assembling Attention Block (LFA) Design. Inspired by [4,18], where low-level feature maps assist on refining decoder features, we explore efficient low-level feature assembling using a transformer block. It can be naturally extended from the transformer block in the encoder without exhausting designs.…”

“…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.…”

“…It is an ill-posed problem because there are 7 unknowns given 3 equations. With recent success of deep learning, deep matting methods [4,11,14,18,22,28,34] achieve promising results on benchmarks such as Composition-1k [34] and alphamatting.com [25]. While increasingly higher accuracy have been promised on benchmarks, due to the limited training/test data, robustness of these methods is still under explored.…”