IP Library Granted Patent US 12,494,040
Granted Patent B2
US 12,494,040 · App. 17/538,555 · Granted Dec 9, 2025

Computer system of unsupervised learning with deep similarity for optical flow estimation and method thereof

Inventors: Sung-Eui Yoon (Daejeon, KR); Woobin Im (Daejeon, KR); Tae Kyun Kim (Daejeon, KR)
Assignee: Korea Advanced Institute Of Science And Technology
G06V10/761G06N20/00G06V10/80G06V10/82H04N19/513G06N3/04G06N3/088G06T7/20G06T2207/10016
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,494,040
App. No.
17/538,555
Granted
Dec 9, 2025
Kind
B2
Abstract

A computer system of unsupervised learning with deep similarity for optical flow estimation is provided. The computer system is configured to calculate deep similarity by using deep features extracted from a sequence of a plurality of images, and learning optical flow for the images based on the deep similarity.

Claims (45)

1 . A method performed by of a computer system, wherein the computer system comprises at least one processor configured to execute computer-readable instructions included in a memory, the method comprising:

calculating, by the at least one processor, deep similarity between locations mapped to each other on a sequence of a plurality of images based on an inner product of deep features of the locations extracted from the sequence of a plurality of images;

generating, by the at least one processor, a deep similarity map in which all locations of the plurality of images are matched respectively based on the deep similarity;

learning, by the at least one processor, optical flow for the plurality of images based on the deep similarity to learn a deep learning model to estimate optical flow of a sequence of input images; and

applying, by the at least one processor, the learned deep learning model to the sequence of input images to estimate the optical flow of the sequence of input images,

wherein the learning of the optical flow comprises:

estimating occlusion locations and non-occlusion locations on the deep similarity map;

determining a similarity threshold value based on first average similarity for the occlusion locations and second average similarity for the non-occlusion locations; and

learning the optical flow by using a loss function having the similarity threshold value.

2 . The method of claim 1 , wherein the deep features are extracted from the plurality of images respectively, by using a deep network consisting of a plurality of layers.

3 . The method of claim 2 , wherein the calculating of the deep similarity comprises:

calculating the deep similarity between the locations mapped to each other on the plurality of images based on the inner product of the deep features of the locations, for each of the layers; and

calculating the deep similarity in the locations by fusing similarities of the layers by a product.

4 . The method of claim 1 , wherein the similarity threshold value is determined as an average value of the first average similarity and the second average similarity.

5 . The method of claim 1 , wherein the loss function is a feature separation loss function configured to increase contrast in the deep similarity map by determining the similarity threshold value based on the first average similarity and the second average similarity in the deep similarity map, and applying the similarity threshold value to amplify a difference between the occlusion locations and the non-occlusion locations.

6 . The method of claim 5 , wherein the learning of the optical flow by using the loss function learns the optical flow by further using at least one of a photometric loss function based on the deep similarity map or a smoothness loss function based on the deep similarity map.

7 . The method of claim 6 , wherein the learning of the optical flow by using the loss function applies different weight values for at least one of the feature separation loss function, the photometric loss function, or the smoothness loss function.

8 . A computer system, comprising:

a memory; and

a processor connected to the memory, and configured to execute at least one instruction stored in the memory, wherein the processor is configured to execute the at least one instruction to:

calculate deep similarity between locations mapped to each other on a sequence of a plurality of images based on an inner product of deep features of the locations extracted from the plurality of images;

generate a deep similarity map in which all locations of the plurality of images are matched respectively based on the deep similarity;

learn optical flow for the plurality of images based on the deep similarity to learn a deep learning model to estimate optical flow of a sequence of input images; and

apply the learned deep learning model to the sequence of input images to estimate the optical flow of the sequence of input images,

wherein the processor is configured to execute the at least one instruction to

estimate occlusion locations and non-occlusion locations on the deep similarity map,

determine a similarity threshold value based on first average similarity for the occlusion locations and second average similarity for the non-occlusion locations, and

learn the optical flow by using a loss function having the similarity threshold value.

9 . The computer system of claim 8 , wherein the deep features are extracted from the plurality of images respectively by using a deep network consisting of a plurality of layers.

10 . The computer system of claim 9 , wherein the processor is configured to execute the at least one instruction to:

calculate the deep similarity between the locations mapped to each other on the images based on the inner product of the deep features of the locations, for each of the layers; and

calculate the deep similarity in the locations by fusing similarities of the layers as by a product.

11 . The computer system of claim 8 , wherein the similarity threshold value is determined as an average value of the first average similarity and the second average similarity.

12 . The computer system of claim 8 , wherein the loss function is a feature separation loss function configured to increase contrast in the deep similarity map calculating the similarity threshold value based on the first average similarity and the second average similarity in the deep similarity map, and applying the similarity threshold value to amplify a difference between the occlusion locations and the non-occlusion locations.

13 . The computer system of claim 12 , wherein the processor is configured to execute the at least one instruction to learn the optical flow by further using at least one of a photometric loss function based on the deep similarity map or a smoothness loss function based on the deep similarity map.

14 . The computer system of claim 13 , wherein the processor is configured to execute the at least one instruction to apply different weight values for at least one of the feature separation loss function, the photometric loss function, or the smoothness loss function.

15 . A non-transitory computer-readable medium for storing at least one program to be executed by at least one processor of a computer system, wherein the at least one processor executes the at least one program to perform a method comprising:

calculating, by the at least one processor deep similarity between locations mapped to each on a sequence of a plurality of images based on an inner product of deep features of the locations extracted from the plurality of images;

generating, by the at least one processor, a deep similarity map in which all locations of the plurality of images are matched respectively based on the deep similarity;

learning, by the at least one processor, optical flow for the plurality of images based on the deep similarity to learn a deep learning model to estimate optical flow of a sequence of input images; and

applying, by the at least one processor, the learned deep learning model to the sequence of input images to estimate the optical flow of the sequence of input images,

wherein the learning of the optical flow comprises:

estimating occlusion locations and non-occlusion locations on the deep similarity map;

determining a similarity threshold value based on first average similarity for the occlusion locations and second average similarity for the non-occlusion locations; and

learning the optical flow by using a loss function having the similarity threshold value.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Nov 30, 2021
From: YOON, SUNG-EUI; IM, WOOBIN; KIM, TAE KYUN
To: KOREA ADVANCED INSTITUTE OF SCIENCE AND TECHNOLOGY
Reel/Frame 058291/0248 →
Priority Claims (1)
KR 10-2021-0062354 · May 14, 2021 · national
Continuity (1)
Related Publication 20220366669A1 · Nov 17, 2022
References Cited (14)
US 10958869B1 · Chi · 2021 [cited by examiner]
US 11790550B2 · Xiao · 2023 [cited by examiner]
US 20200084427A1 · Sun · 2020 [cited by examiner]
US 20200211206A1 · Wang · 2020 [cited by examiner]
US 20200265590A1 · Daniilidis · 2020 [cited by examiner]
US 20210142497A1 · Pugh · 2021 [cited by examiner]
US 20220156946A1 · Lin · 2022 [cited by examiner]
US 20220335624A1 · Maurer · 2022 [cited by examiner]
CN 108805900B · 2021 [cited by examiner]
KR 1020200010971 · 2020 [cited by applicant]
Meister, Simon, Junhwa Hur, and Stefan Roth. “Unflow: Unsupervised learning of optical flow with a bidirectional census loss.” Proceedings of the AAAI conference on artificial intelligence. vol. 32. No. 1. pp. 7251-7259… [cited by examiner]
Li, Jianfeng, et al. “Occlusion aware unsupervised learning of optical flow from video.” Thirteenth International Conference on Machine Vision. vol. 11605. SPIE, Jan. 4, 2021. pp. 1-8 (Year: 2021). [cited by examiner]
Im, W et al., “Unsupervised Learning of Optical Flow with Deep Feature Similarity,” 16 [cited by applicant]
Teed Z. et al., RAFT: Recurrent All-Pairs Field Transforms for Optical Flow; Princeton University, Aug. 25, 2020, 21 pp. [cited by applicant]