Amodal Segmentation through Out-of-Task and Out-of-Distribution with a Bayesian Model#

  • Authors: Yihong Sun, Adam Kortylewski, Alan Yuille

  • Affiliations: Johns Hopkins University

  • CVPR, 2022

  • Links: project

Summary#

Amodal segmentation aims to segment object boundaries which are occluded and hence invisible. The authors formulate amodal segmentation as an out-of-task and out-of-distribution generalization problem. The model is trained from non-occluded images using bounding box annotations and class labels only, but is applied to generalize out-of-task to object segmentation and to generalize out-of-distribution to segment occluded objects. The proposed approach outperfoms alternate methods with the same supervision by a wide margin, and even outperforms supervised methods when the occlusion is large.

Key Ideas#

The model taks as input a feature map \(\bar{F} = \psi(I, \eta)\) where \(I\) is the input image and \(\zeta\) are the neural network weights. Let the features within a given bounding box \(\mathcal{D}\) be \(F = \{ f_a: a \in \mathcal{D} \}\).

A Bayesian model for amodal segmentation. The authors introduce a latent variable \(w_a \in \{0, 1\}\) to indicate foreground/background which are learnt without additional supervision.

\[P_a(f_a \mid y, m, w_a) = P_a(f_a \mid y, m)^{w_a} B_a(f_a \mid y, m)^{1-w_a} \times P_a(w_a \mid y, m)\]

The foreground and background model \(P_a(f_a \mid y, m), B_a(f_a \mid y, m)\) are mixtures of von Mises Fisher distributions.

Shape modeling. The authors also introduce shape priors \(P(\overrightarrow{w} \mid y, m) = \prod_{a \in \mathcal{D} P_a(w_a \mid y, m)\), a learned 2D spatial map conditioned on the object category \(y\) and the class mixture \(m\). Finally this gives a generative model of this data:

\[P(F \mid y) = \sum_{m, \overrightarrow{w}} P(F \mid y, m, \overrightarrow{w}) P(m) P(\overrightarrow{w} \mid y, m)\]

Technical Details#

Notes#

References#

[1] Y. Sun, A. Kortylewski, A. Yuille. “Amodal segmentation through out-of-task and out-of-distribution with a bayesian model.”. In CVPR, 2021.