Human Instance Matting via Mutual Guidance and Multi-Instance Refinement

Fast development of mobile internet technology has triggered the rapid growth of multimedia industry especially we-media, where users are heavily engaged in editing tools to beautify or re-create their image and video contents. As one of the primary techniques for efficient image editing, image matting has achieved significant improvement with the wide adoption of deep neural networks in the task. However, existing matting methods still fail or else are not easy to use in many scenarios, such as extracting the foreground human while removing background humans, or instance-level editing as shown in Figure 1: what if we want to independently extract and edit each human instance?

Similar to semantic versus instance segmentation, existing matting methods, which focus on a region based on a given trimap or a known object class, are unable to differentiate instances. To address this issue, we propose a new task called human instance matting (HIM), which aims to automatically extract precise alpha matte for each human instance in a given image. HIM shares similarities to the following conventional tasks while embodying fundamental differences making it a problem on its own: 1) instance segmentation aims at distinguishing instances, but it can only produce sharp object boundary without semi-transparency consideration; 2) recent soft segmentation [3] is capable of generating soft segments for multiple instances of different classes with instance-aware features, but cannot deal with instances of the same class; 3) conventional matting aims at extracting precise alpha matte, but it lacks instance awareness. Overall, human instance matting is a unified task encompassing the characteristics of the aforementioned related tasks while introducing new technical challenges.

Conventional matting is based on the image compositing equation where an image $I$ is the combination of foreground $F$ layer, background $B$ layer modulated by alpha $\alpha$:

$$I=\alpha F+(1-\alpha)B.$$

(1)

To adapt to multiple instance matting, we modify the 2-layer Equation 1 to one of multi-instance layered composition, where each instance layer is attenuated by its corresponding $\alpha$:

$$I=\sum_{i=0}^{n}\alpha_{i}L_{i},\hskip 14.45377pt\textrm{s.t.}\sum_{i=0}^{n}\alpha_{i}=1$$

(2)

where $L_{i}$ and $\alpha_{i}$ respectively denote the foreground and alpha matte for instance $i>0$; $L_{0}$ and $\alpha_{0}$ respectively represent the background and its corresponding alpha matte; $n$ is the number of instances. This equation had also appeared in [34, 3], but all such relevant matting and segmentation tasks were not instance aware. The goal of instance matting is to solve for target mattes $\alpha_{i}$ for all $i>0$.

By exploring the complex relation among multiple instances, we propose a new instance matting framework, called InstMatt, where a novel mutual guidance strategy enables a deep model to decompose mingled compositing colors into their respective instances. Our mutual guidance strategy takes both the relation between instances and the background, and the relation among instances into consideration. Besides, a multi-instance refinement module is carefully designed and engineered for interchanging information among instances to synchronize predictions for further refinement. Equipped with the novel mutual guidance and multi-instance refinement, our InstMatt is able to not only produce high-quality human alpha matte but also distinguish multiple human instances shown in Figure 1.

With this new HIM task, existing evaluation metrics for instance segmentation or matting are insufficient, which were designed for either one of the tasks. We propose a new metric, called instance matting quality (IMQ), that simultaneously measures instance recognition quality and alpha matte quality. To provide a general and comprehensive validation on instance matting techniques, we construct an instance matting benchmark, HIM2K, which consists of a synthetic image benchmark and a natural image benchmark totaling 2,000 images with high-quality matte ground truths.

To demonstrate the promise of our technical contributions beyond human instance matting, we present preliminary results on matting multi-object instances not limited to humans, a fruitful future direction to explore.

Our HIM framework, called InstMatt, consists of two steps, first recognizing instances and then extracting their respective alpha mattes. This allows the model to globally discover instances and then refine them according to local context. Figure 2 illustrates the whole framework.

Existing benchmarks are designed for instance segmentation such as COCO dataset [40], or matting such as Composition-1K [60], but not for instance matting. They cannot provide a comprehensive evaluation for instance matting. In this paper, we propose a human instance matting benchmark called HIM2K, which is composed of two subsets, synthetic image subset and natural image subset respectively containing 1,680 and 320 images.

If we use layer $L$ to represent a foreground image $F$ or background image $I_{0}$, Equation 14 for the last iteration can be simplified as Equation 15 which is the same as Equation 2:

$$I=\sum_{i=0}^{n}\alpha_{i}^{{}^{\prime}}L_{i}$$

(15)

where $\alpha_{i}^{{}^{\prime}}$ denotes the alpha matte of $i$-th layer $L_{i}$, the target to be estimated for instance $i$ when $i>0$.

In this section, we introduce a new metric for instance matting. Existing metrics are designed for either matting or instance segmentation including semantic segmentation. Instance segmentation metrics, such as mask average precision (mask AP), are used for measuring the binary instance mask quality, and thus unsuitable for evaluating alpha matte with fractional values in transitional region. On the other hand, the most widely used matting metrics, namely, the four errors MAD (or SAD), MSE, Gradient and Connectivity, measure alpha matte quality without instance awareness. The above limitations of existing metrics necessitate a new metric, which we call instance matting quality (IMQ).

In this section, we introduce our synthetic training dataset, evaluation and ablation studies. More details about the implementation including the network structure, data augmentations and training schedule can be found in the supplementary materials.

In this paper, we propose a new task, instance matting with human instance matting as the first significant example by proposing a novel instance matting framework. Our InstMatt utilizes mutual exclusive guidance to guide the matting branch to extract alpha matte for each instance, which is followed by a multi-instance refinement module to synchronize information among co-occurring instances. InstMatt is capable of handling challenging cases with multiple and overlapping instances, which can be adapted to other semantic class instance matting beyond human instances. We hope the proposed method, alongside with the new instance matting metric and the human instance matting benchmark, will encourage more future works.

From the perspective of functionality, our InstMatt consists of two steps, that is, instance recognition and mask refinement. We adopt MaskRCNN [23] with the backbone ResNet50 [25] as our instance recognition model since MaskRCNN is a conventional and also competitive approach in instance segmentation. We take the publicly released MaskRCNN pre-trained weight from Detectron2 [59] without further finetuning because this model is well-trained on the large-scale COCO [40] dataset containing rich scenarios which can be well generalized to other datasets.

The mask refinement step in InstMatt can be further divided into two modules, i.e., tri-mask guided matting branch and multi-instance refinement. We adopt the network used in MG [61] as our matting branch. The network takes ResNet34 [25] as the backbone and applies three convolution blocks with a stride of $8$, $4$ and $1$ in the decoder respectively to reconstruct the features for tri-matte prediction. During inference stage, after extracting the instance masks, we obtain a tri-matte for each instance. Next, all the tri-mattes are sent to the multi-instance refinement for information synchronization.

In implementation, the multi-instance refinement contains 4 learnable layers as illustrated in Figure 10. The four convolution layers all utilize $3\times 3$ kernel.

During training, we use two datasets, one of which is the synthetic training dataset mentioned in the main paper Section 6.1. Since the synthetic training samples have a domain gap with natural images, we also include a natural set containing 41330 images selected from the COCO training set. However, it is non-trivial to adapt the COCO dataset to the instance matting task due to the lack of instance matting annotations. To tackle this issue, we adopt a partial supervision strategy to make the samples with segmentation annotations applicable in our task.

To enrich the training dataset and avoid overfitting, various augmentation operations are adopted on the training samples. Besides random flip, random zoom, random shearing, as well as random crop, we propose tri-mask augmentation to improve the fault tolerance of the model, in particular, robustness against missing instances and imperfect masks.

In the first step, for an image with $n$ instances, we adopt two ways to generate masks for these instances. For a subset of the $n$ instances, we obtain their masks from the instance segmentation model; for the rest of the instances, we generate their masks from a random truncation on the ground truth alpha matte. A hybrid of the two ways in instance generation increases the diversity of masks.

In the second step, we first randomly pick an instance $i$ as the target instance, then randomly choose a subset from the rest of $n-1$ instances to produce $M_{r}$. Finally we obtain $M_{b}$ by $1-M_{t}\cup M_{r}$. Such relaxation operation on $M_{r}$ make the model ascribe the pixels of those undetected instances to $\alpha_{r}$, rather than $\alpha_{t}$ or $\alpha_{b}$.

In the last step, we randomly dilate or erode or perform a hybrid of dilation and erosion on the tri-mask with kernel size in $[1,30]$. Such perturbation on tri-mask further improves the fault tolerance of our model.

Note that the ground truth tri-matte for the tri-mask are generated without the relaxation or perturbation operation.

Their generation follows Equation 21–23:

	$\displaystyle\alpha_{i,t}$	$\displaystyle=\alpha_{i}$		(21)
	$\displaystyle\alpha_{i,r}$	$\displaystyle=\sum\limits_{j=1\ \text{and}\ j\neq i}^{n}\alpha_{j}$		(22)
	$\displaystyle\alpha_{i,b}$	$\displaystyle=1-\alpha_{i,t}-\alpha_{i,r}$		(23)

Without and with the tri-mask augmentation, the IMQ_mad of our InstMatt is 67.51 and 69.40 as tabulated in Table 4, showing a promotion benefiting from the tri-mask augmentation.

Our training schedule consists of two steps:

1. 1.

   Train the matting branch on the synthetic and natural training datasets. The branch is initialized with ImageNet [17] pre-trained weight. We use a batch size of 16 in total on 4 GPU cards. Adam optimizer with $\beta_{1}=0.5$ and $\beta_{2}=0.999$ is adopted. The initial learning rate is set to $0.001$ and decays at a cosine learning rate [45, 21]. The training lasts for 100,000 iterations with a warm-up of the first 5, 000 iterations.

2. 2.

   After the matting branch is well-trained, we freeze the matting branch and train the multi-instance refinement module. In the second step, we use a batch size of 4 in total on 4 GPU cards. The initial learning rate is set to $0.0001$. The training lasts for 25,000 iterations with a warm-up of the first 1, 000 iterations. We keep the other hyper parameters the same as those in the first step.

During similarity measurement, we adopt the widely used error functions in conventional matting task to evaluate the instance alpha matte from multiple dimensions. In implementation, we compute $\mathcal{E}(\alpha,\hat{\alpha})$ as follows,

	$\displaystyle\mathcal{E}(\alpha,\hat{\alpha})$	$\displaystyle=\frac{1}{|\mathcal{P}|}\sum_{p\in\mathcal{P}}\mathcal{E}(\alpha_{p},\hat{\alpha}_{p})$		(24)
	$\displaystyle\mathcal{P}$	$\displaystyle=[\alpha>0]\cup[\hat{\alpha}>0]$		(25)

We take the average upon the union of quantified $\alpha$ and $\hat{\alpha}$ instead of the whole image to avoid the overwhelming zero values from the large amount of background pixels especially for small instances.

On the contrary, besides refining the instance alpha matte along the boundary and the hairy regions among the TP set, our InstMatt is also capable of recognizing an instance and correspondingly extracting its alpha matte even though only a low-quality mask with misleading instance information is provided, such as the first example in Figure 5 in the main paper and the example in Figure 13.

Figure 14 shows the cases containing multiple instances next to each other closely. MaskRCNN is able to distinguish instances but produces overlapping instance masks, which cannot be addressed either by CascadePSP [12] or a naive extension of existing matting models [36, 52, 18, 61]. Our InstMatt, however, can clearly separate the instances and generate non-overlapping instance alpha mattes.

Figure 15 shows the cases with occlusion. Under such cases, a part of one instance, usually a hand, or an arm, appears within the region of another instance and is far away from its own body. It is difficult to solve these cases for recognition tasks due to the limitation of receptive fields and the bottleneck of long-range feature propagation. As shown in Figure 15, both instance segmentation models and matting models fall short of producing satisfactory results in these cases, while our InstMatt still produces promising results. Inter-instance mutual exclusive information guides the model to retrieve the remote pixels sharing the similar appearance with the body region instead of ascribing them to the other instances.

Figure 16 compares the performance on a case with incomplete instance segmentation masks. Although CascadePSP or matting-based models are able to refine the mask along the boundary, they cannot recover a part of missing region due to the lack of instance awareness. Our InstMatt can find the lost region from the background giving the credit to the mutual exclusive guidance between the human instances and the background.

Figure 17 compares the performance on small instances. Compared to other methods, our InstMatt still shows stable performance on the instances of small or tiny scales, demonstrating the generalization ability and the fineness of our results.

Under most cases, our method is capable of upgrading a low-quality instance mask into a high-quality alpha matte. However, sometimes the instance segmentation model cannot differentiate two largely overlapping instances as shown in Figure 18. Note that this example is different from the one in Figure 13. In Figure 13, MaskRCNN recognizes two instances although the mask of instance 2 covers a part of instance 1. Differently, in Figure 18, the instance segmentation model regards two left human instances as one and only predict a mask for the two left instances. In this case, our model can only refine the mask but cannot separate them due to the lack of sufficient guidance.

Examples with hair details in alpha format are provided in Figure 19. More qualitative comparisons on images containing complex multiple overlapping or crowded cases are shown in Figure 20.