# US 2026/0311469 A1 · Instance segmentation method and system based on self-error estimation, and method and system for training mask estimation model therefor

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2026/0311469 A1
- Document ID: US20260311469A1
- Status: Published application
- Kind: A1
- Application number: 19/459,284
- Filed: 2026-01-26
- Published: 2026-10-08
- Earliest claimed: 2025-04-02
- Assignees: Gwangju Institute of Science and Technology
- Applicants: Gwangju Institute of Science and Technology
- Inventors: Kyoobin LEE; Seung Hyeok BACK; Sang Beom LEE; Kang Min KIM; Sung Ho SHIN
- Web page: https://patentgrep.com/patent/US20260311469A1

## CPC

- G06V 10/776: Validation; Performance evaluation
- G06V 10/25: Determination of region of interest [ROI] or a volume of interest [VOI]
- G06V 10/267: {by performing operations on regions, e.g. growing, shrinking or watersheds}
- G06V 10/82: using neural networks
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING

## Abstract

An instance segmentation method including: receiving a target image; estimating an initial bounding box corresponding to a specific object from the target image and estimating, based on the initial bounding box, an initial instance mask for the specific object; predicting a self-error for the initial instance mask based on the initial bounding box; and estimating, from the target image, a bounding box corrected according to the self-error, in consideration of the self-error for the initial bounding box and estimating an instance mask, based on the corrected bounding box and the self-error.

## Claims

### Claim 1

An instance segmentation method, comprising:
receiving a target image;
estimating an initial bounding box corresponding to a specific object from the target image and estimating, based on the initial bounding box, an initial instance mask for the specific object;
predicting a self-error for the initial instance mask based on the initial bounding box; and
estimating, from the target image, a bounding box corrected according to the self-error, in consideration of the self-error for the initial bounding box and estimating an instance mask, based on the corrected bounding box and the self-error.

### Claim 2 (depends on 1)

The instance segmentation method of claim 1, wherein the estimating of the initial instance mask comprises:
estimating, by using a pre-trained mask estimation model, a region corresponding to the specific object from the target image to generate the initial bounding box; and
estimating pixels corresponding to the specific object from the initial bounding box to generate the initial instance mask.

### Claim 3 (depends on 2)

The instance segmentation method of claim 2, wherein the predicting of the self-error comprises:
combining, through the mask estimation model, in the target image, a region of interest according to the initial bounding box and the initial instance mask, and predicting the self-error from combined data.

### Claim 4 (depends on 3)

The instance segmentation method of claim 3, wherein the estimating of the instance mask based on the self-error comprises:
combining, through the mask estimation model, the predicted self-error to the initial bounding box to estimate a bounding box in which the initial bounding box is corrected according to the self-error; and
combining the corrected bounding box corrected according to the self-error and the self-error, and estimating the instance mask corrected according to the self-error from combined data.

### Claim 5 (depends on 2)

The instance segmentation method of claim 2, wherein the mask estimation model, as a single model, is implemented based on a configuration of Mask R-CNN.

### Claim 6 (depends on 1)

The instance segmentation method of claim 1, wherein the self-error is predicted as a pixel-wise error map having a same spatial resolution as the initial instance mask.

### Claim 7 (depends on 6)

The instance segmentation method of claim 6, wherein the self-error comprises a plurality of pixel-wise probability values indicating whether each pixel of the initial instance mask is correctly segmented or erroneously segmented.

### Claim 8 (depends on 7)

The instance segmentation method of claim 7, wherein the self-error is represented as a multi-channel error map including at least a first channel corresponding to correctly segmented pixels and a second channel corresponding to erroneously segmented pixels.

### Claim 9 (depends on 1)

The instance segmentation method of claim 1, wherein the estimating of the corrected bounding box comprises calculating correction parameters including at least one of a positional offset and a scale adjustment derived from the predicted self-error.

### Claim 10 (depends on 9)

The instance segmentation method of claim 9, wherein the corrected bounding box is estimated by applying the correction parameters to the initial bounding box using a predefined mathematical transformation.

### Claim 11 (depends on 1)

The instance segmentation method of claim 1, wherein the estimating of the corrected bounding box and the estimating of the instance mask are repeatedly performed a plurality of times based on the self-error.

### Claim 12 (depends on 11)

The instance segmentation method of claim 11, wherein the repeated estimating is terminated when a change in the self-error between successive iterations is less than a predetermined threshold.

### Claim 13 (depends on 2)

The instance segmentation method of claim 2, wherein the mask estimation model further includes a self-error prediction head configured to predict the self-error jointly with bounding box estimation and instance mask estimation.

### Claim 14 (depends on 13)

The instance segmentation method of claim 13, wherein the self-error prediction head is configured to receive, as input, a combination of region-of-interest features corresponding to the initial bounding box and features derived from the initial instance mask.

### Claim 15 (depends on 4)

The instance segmentation method of claim 4, wherein the corrected instance mask is estimated by combining the corrected bounding box and the self-error as input features to the mask estimation model.

### Claim 16 (depends on 1)

The instance segmentation method of claim 1, wherein the mask estimation model is trained to minimize a loss function including at least one of a mask loss, a bounding box loss, and a self-error loss.

### Claim 17 (depends on 16)

The instance segmentation method of claim 16, wherein the self-error loss is calculated based on a comparison between a predicted self-error and a ground-truth self-error derived from a ground-truth instance mask.

### Claim 18 (depends on 5)

The instance segmentation method of claim 5, wherein the mask estimation model is implemented as a single unified network in which a bounding box head, a mask head, and a self-error prediction head share at least a portion of feature extraction layers.

### Claim 19

An instance segmentation system, comprising:
a storage unit in which a target image is stored; and
a control unit configured to estimate an instance mask from the target image, wherein the control unit:
estimates an initial bounding box corresponding to a specific object from the target image;
estimates, based on the initial bounding box, an initial instance mask for the specific object;
predicts a self-error for the initial instance mask based on the initial bounding box;
estimates, from the target image, a bounding box corrected according to the self-error, in consideration of the self-error for the initial bounding box; and
estimates an instance mask, based on the corrected bounding box and the self-error.

### Claim 20

A method for training a mask estimation model, comprising:
receiving training images and ground-truth masks which are labeling data for the training images;
inputting the training images to a mask estimation model to estimate a training initial instance mask for the training images;
predicting a self-error for the training initial instance mask through the mask estimation model, and estimating, based on the self-error, a training instance mask corrected from the training initial instance mask; and
comparing the training instance mask and the ground-truth masks, and training, based on a comparison result, the mask estimation model.
