Throughout my time at Berkeley, I have engaged in close collaborations with Dr. Yubei Chen at
Meta AI Research / UC Davis.
I also worked closely with Professor Stella Yu during the initial two years of my PhD.
Prior to joining Berkeley, I had the privilege of working with Dr. Tyng-Luh Liu at IIS, Academia Sinica.
I also spent half a year as a Visiting Researcher at UC Berkeley / ICSI from 2018 to 2019.
I am passionate about building universal models that integrate source from various modalities,
with a particular focus on aligning vision with language.
I am also interested in the area of self-supervised representation learning and understanding,
with an emphasis on its application to image and video tasks.
UC Berkeley Ph.D. Student Sept. 21 - Present
Adobe Inc. Research Intern May. 24 - Present May. 22 - Nov. 22
IIS, Academia Sinica Research Assistant Apr. 20 - Aug. 21
Accurate diagnosis of ocular surface diseases is critical in optometry and ophthalmology,
which hinge on integrating clinical data sources (e.g., meibography imaging and clinical metadata).
Traditional human assessments lack precision in quantifying clinical observations, while current machine-based
methods often treat diagnoses as multi-class classification problems, limiting the diagnoses to a predeļ¬ned
closed-set of curated answers without reasoning the clinical relevance of each variable to the diagnosis.
To tackle these challenges, we introduce an innovative multi-modal diagnostic pipeline (MDPipe) by employing
large language models (LLMs) for ocular surface disease diagnosis.
We first employ a visual translator to interpret meibography images by converting them into quantifiable
morphology data, facilitating their integration with clinical metadata and enabling the communication of
nuanced medical insight to LLMs. To further advance this communication, we introduce a LLM-based summarizer
to contextualize the insight from the combined morphology and clinical metadata, and generate clinical report
summaries. Finally, we refine the LLMs' reasoning ability with domain-specific insight from real-life clinician
diagnoses. Our evaluation across diverse ocular surface disease diagnosis benchmarks demonstrates that MDPipe
outperforms existing standards, including GPT-4, and provides clinically sound rationales for diagnoses.
@inproceedings{yeh2024insight,
title={Insight: A Multi-modal Diagnostic
Pipeline Using LLMs for Ocular Surface
Disease Diagnosis},author={Yeh, Chun-Hsiao
and Wang, Jiayun and Graham, Andrew D and
Liu, Andrea J and Tan, Bo and Chen, Yubei
and Ma, Yi and Lin, Meng C},booktitle=
{International Conference on Medical Image
Computing and Computer-Assisted
Intervention},pages={711--721},year={2024},
organization={Springer}
}
Recent text-to-image diffusion models are able to learn and synthesize images containing novel, personalized concepts (e.g.,
their own pets or specific items) with just a few examples for training. This paper tackles two interconnected issues within
this realm of personalizing text-to-image diffusion models. First, current personalization techniques fail to reliably extend
to multiple concepts --- we hypothesize this to be due to the mismatch between complex scenes and simple text descriptions in
the pre-training dataset (e.g., LAION). Second, given an image containing multiple personalized concepts, there lacks a holistic
metric that evaluates performance on not just the degree of resemblance of personalized concepts, but also whether all concepts
are present in the image and whether the image accurately reflects the overall text description. To address these issues,
we introduce Gen4Gen, a semi-automated dataset creation pipeline utilizing generative models to combine personalized concepts
into complex compositions along with text-descriptions. Using this, we create a dataset called MyCanvas, that can be used to
benchmark the task of multi-concept personalization. In addition, we design a comprehensive metric comprising two scores
(CP-CLIP and TI-CLIP) for better quantifying the performance of multi-concept, personalized text-to-image diffusion methods.
We provide a simple baseline built on top of Custom Diffusion with empirical prompting strategies for future researchers to
evaluate on MyCanvas. We show that by improving data quality and prompting strategies, we can significantly increase
multi-concept personalized image generation quality, without requiring any modifications to model architecture or training algorithms.
@article{yeh2024gen4gen,
title={Gen4Gen: Generative Data
Pipeline for Generative Multi-Concept
Composition},author={Yeh, Chun-Hsiao and
Cheng, Ta-Ying and Hsieh, He-Yen and
Lin, Chuan-En and Ma, Yi and Markham,
Andrew and Trigoni, Niki and
Kung, Hsiang-Tsung and Chen, Yubei},
journal={arXiv preprint arXiv:2402.15504},
year={2024}
}
Creating content for a specific identity (ID) has shown significant interest in the field of generative models. In the field of
text-to-image generation (T2I), subject-driven content generation has achieved great progress with the ID in the images controllable.
However, extending it to video generation is not well explored. In this work, we propose a simple yet effective subject identity controllable
video generation framework, termed Video Custom Diffusion (VCD). With a specified subject ID defined by a few images, VCD reinforces the
identity information extraction and injects frame-wise correlation at the initialization stage for stable video outputs with identity
preserved to a large extent. To achieve this, we propose three novel components that are essential for high-quality ID preservation:
1) an ID module trained with the cropped identity by prompt-to-segmentation to disentangle the ID information and the background noise
for more accurate ID token learning; 2) a text-to-video (T2V) VCD module with 3D Gaussian Noise Prior for better inter-frame consistency
and 3) video-to-video (V2V) Face VCD and Tiled VCD modules to deblur the face and upscale the video for higher resolution.
Despite its simplicity, we conducted extensive experiments to verify that VCD is able to generate stable and high-quality videos with
better ID over the selected strong baselines. Besides, due to the transferability of the ID module, VCD is also working well with
finetuned text-to-image models available publically, further improving its usability.
@article{ma2024magic,
title={Magic-Me:
Identity-Specific Video
Customized Diffusion},
author={Ma, Ze and Zhou, Daquan
and Yeh, Chun-Hsiao and
Wang, Xue-She and Li, Xiuyu
and Yang, Huanrui and Dong, Zhen
and Keutzer, Kurt and Feng, Jiashi},
journal={arXiv preprint arXiv:2402.09368},
year={2024}
}
Large-scale vision-language models (VLM) have shown impressive results for language-guided search applications.
While these models allow category-level queries, they currently struggle with personalized searches for moments
in a video where a specific object instance such as ``My dog Biscuit'' appears.
We present the following three contributions to address this problem.
First, we describe a method to meta-personalize a pre-trained VLM, learning how to learn to personalize a VLM at test time to search in video.
Our method extends the VLM's token vocabulary by learning novel word embeddings specific to each instance.
To capture only instance-specific features, we represent each instance embedding as a combination of shared and learned global category features.
Second, we propose to learn such personalization without explicit human supervision.
Our approach automatically identifies moments of named visual instances in video using transcripts and vision-language similarity in the VLM's embedding space.
Finally, we introduce This-Is-My, a personal video instance retrieval benchmark.
We evaluate our approach on This-Is-My and DeepFashion2 and show that we obtain a 15% relative improvement over the state of the art on the latter dataset.
@inproceedings{yeh2023meta,
title={Meta-Personalizing
Vision-Language Models To Find Named
Instances in Video},
author={Yeh, Chun-Hsiao and Russell, Bryan
and Sivic, Josef and Heilbron, Fabian Caba
and Jenni, Simon},
booktitle={Proceedings of the IEEE/CVF
Conference on Computer Vision and
Pattern Recognition},
pages={19123--19132},
year={2023}
}
Decoupled Contrastive Learning Chun-Hsiao Yeh, Cheng-Yao Hong, Yen-Chi Hsu, Tyng-Luh Liu, Yubei Chen, and Yann LeCun
European Conference on Computer Vision (ECCV), 2022.
Contrastive learning (CL) is one of the most successful paradigms for self-supervised learning (SSL).
In a principled way, it considers two augmented "views" of the same image as positive to be pulled closer,
and all other images as negative to be pushed further apart. However, behind the impressive success of CL-based techniques,
their formulation often relies on heavy-computation settings, including large sample batches, extensive training epochs, etc.
We are thus motivated to tackle these issues and establish a simple, efficient, yet competitive baseline of contrastive learning.
Specifically, we identify, from theoretical and empirical studies, a noticeable negative-positive-coupling (NPC) effect in the
widely used InfoNCE loss, leading to unsuitable learning efficiency concerning the batch size. By removing the NPC effect,
we propose decoupled contrastive learning (DCL) loss, which removes the positive term from the denominator and significantly
improves the learning efficiency. DCL achieves competitive performance with less sensitivity to sub-optimal hyperparameters,
requiring neither large batches in SimCLR, momentum encoding in MoCo, or large epochs. We demonstrate with various benchmarks
while manifesting robustness as much less sensitive to suboptimal hyperparameters. Notably, SimCLR with DCL achieves 68.2%
ImageNet-1K top-1 accuracy using batch size 256 within 200 epochs pre-training, outperforming its SimCLR baseline by 6.4%. Further,
DCL can be combined with the SOTA contrastive learning method, NNCLR, to achieve 72.3% ImageNet-1K top-1 accuracy with 512 batch
size in 400 epochs, which represents a new SOTA in contrastive learning. We believe DCL provides a valuable baseline for future
contrastive SSL studies.
@inproceedings{yeh2022decoupled,
title={Decoupled contrastive learning},
author={Yeh, Chun-Hsiao and Hong,
Cheng-Yao and Hsu, Yen-Chi and Liu,
Tyng-Luh and Chen, Yubei and LeCun, Yann},
booktitle={Computer Vision--ECCV 2022:
17th European Conference, Tel Aviv, Israel,
October 23--27, 2022, Proceedings, Part XXVI},
pages={668--684},
year={2022},
organization={Springer}
}
Self-supervised training that elegantly couples contrastive learning with a wide spectrum of data augmentation
techniques has been shown to be a successful paradigm for representation learning. However, current methods
implicitly maximize the agreement between differently augmented views of the same sample, which may perform
poorly in certain situations. For example, considering an image comprising a boat on the sea, one augmented view
is cropped solely from the boat and the other from the sea, whereas linking these two to form a positive pair
could be misleading. To resolve this issue, we introduce a Self-Augmentation with Guided Attention (SAGA) strategy,
which augments input data based on predictive attention to learn representations rather than simply applying
off-the-shelf augmentation schemes. As a result, the proposed self-augmentation framework enables feature learning
to enhance the robustness of representation.
@inproceedings{yeh2022saga,
title={SAGA: Self-Augmentation
with Guided Attention for
Representation Learning},
author={Yeh, Chun-Hsiao and
Hong, Cheng-Yao and Hsu, Yen-Chi
and Liu, Tyng-Luh},
booktitle={ICASSP 2022-2022 IEEE
International Conference on Acoustics,
Speech and Signal Processing (ICASSP)},
pages={3463--3467},
year={2022},
organization={IEEE}
}
Deep learning approaches are widely explored in safety-critical autonomous driving systems
on various tasks. Network models, trained on big data, map input to probable prediction
results. However, it is unclear how to get a measure of confidence on this prediction at
the test time.Our approach to gain this additional information is to estimate how similar
test data is to the training data that the model was trained on. We map training instances
onto a feature space that is the most discriminative among them. We then model the entire
training set as a Gaussian distribution in that feature space. The novelty of the test data
is characterized by its low probability of being in that distribution, or equivalently a
large Mahalanobis distance in the feature space.Our distance metric in the discriminative
feature space achieves a better novelty prediction performance than the state-of-the-art
methods on most classes in CIFAR-10 and ImageNet. Using semantic segmentation as a proxy
task often needed for autonomous driving, we show that our unsupervised novelty prediction
correlates with the performance of a segmentation network trained on full pixel-wise annotations.
These experimental results demonstrate potential applications of our method upon identifying
scene familiarity and quantifying the confidence in autonomous driving actions.
@inproceedings{ranjbar2020scene,
title={Scene novelty prediction from
unsupervised discriminative feature learning},
author={Ranjbar, Arian and Yeh, Chun-Hsiao
and Hornauer, Sascha and Stella, X Yu
and Chan, Ching-Yao},
booktitle={2020 IEEE 23rd International
Conference on Intelligent Transportation
Systems (ITSC)},
pages={1--7},
year={2020},
organization={IEEE}
}
Vulnerability of recognition systems to spoofing attacks (presentation attacks) is
still an open security issue in the biometrics domain. Among all biometric traits,
face is exposed to the most serious threat since it is particularly easy to access
and reproduce. In this paper, an effective approach against face spoofing attacks
based on perceptual image quality assessment features with multiscale analysis is
presented. First, we demonstrate that the recently proposed blind image quality
evaluator (BIQE) is effective in detecting spoofing attacks. Next, we combine
the BIQE with an image quality assessment model called effective pixel similarity
deviation (EPSD), which we propose to obtain the standard deviation of the gradient
magnitude similarity map by selecting effective pixels in the image. A total number
of 21 features acquired from the BIQE and EPSD constitute the multi-scale descriptor
for classification. Extensive experiments based on both intradataset and cross-dataset
protocols were performed using three existing benchmarks, namely, Replay-Attack, CASIA,
and UVAD. The proposed algorithm demonstrated its superiority in detecting face spoofing
attacks over many state of the art methods. We believe that the incorporation of the image
quality assessment knowledge into face liveness detection is promising to improve the overall accuracy.
@inproceedings{yeh2018face,
title={Face liveness detection
based on perceptual image quality
assessment features with multi-scale analysis},
author={Yeh, Chun-Hsiao and Chang, Herng-Hua},
booktitle={2018 IEEE Winter conference
on applications of computer vision (WACV)},
pages={49--56},
year={2018},
organization={IEEE}
}
Face recognition has been extensively used in a wide variety of security systems for
identity authentication for years. However, many security systems are vulnerable to
spoofing face attacks (e.g., 2D printed photo, replayed video). Consequently, a number
of anti-spoofing approaches have been proposed. In this study, we introduce a new
algorithm that addresses the face liveness detection based on the digital focus
technique. The proposed algorithm relies on the property of digital focus with
various depths of field (DOFs) while shooting. Two features of the blurriness
level and the gradient magnitude threshold are computed on the nose and the
cheek subimages. The differences of these two features between the nose and
the cheek in real face images and spoofing face images are used to facilitate
detection. A total of 75 subjects with both real and spoofing face images were
used to evaluate the proposed framework. Preliminary experimental results
indicated that this new face liveness detection system achieved a high
recognition rate of 94.67% and outperformed many state-of-the-art methods.
The computation speed of the proposed algorithm was the fastest among the tested methods.
@inproceedings{yeh2017face,
title={Face liveness detection with
feature discrimination between sharpness
and blurriness},
author={Yeh, Chun-Hsiao and Chang, Herng-Hua},
booktitle={2017 Fifteenth IAPR International
Conference on Machine Vision Applications (MVA)},
pages={398--401},
year={2017},
organization={IEEE}
}
Projects
Comics Generation NTU CSIE ADLxMLDS 2017 Fall Project
Tensorflow implementation of Conditional Generative Adversarial Network (CGAN) automatically
generates anime images based on given constraints (e.g., green hair, blue eyes).
Video Captioning NTU CSIE ADLxMLDS 2017 Fall Project
TensorFlow implementation of Seq2seq model (S2VT) and attention mechanism,
which generates the description (captions) for the given video.