Minchul Shin

Changdae Oh, Junhyuk So, Hoyoon Byun et al.

Pre-trained multi-modal models, such as CLIP, provide transferable embeddings and show promising results in diverse applications. However, the analysis of learned multi-modal embeddings is relatively unexplored, and the embedding transferability can be improved. In this work, we observe that CLIP holds separated embedding subspaces for two different modalities, and then we investigate it through the lens of uniformity-alignment to measure the quality of learned representation. Both theoretically and empirically, we show that CLIP retains poor uniformity and alignment even after fine-tuning. Such a lack of alignment and uniformity might restrict the transferability and robustness of embeddings. To this end, we devise a new fine-tuning method for robust representation equipping better alignment and uniformity. First, we propose a Geodesic Multi-Modal Mixup that mixes the embeddings of image and text to generate hard negative samples on the hypersphere. Then, we fine-tune the model on hard negatives as well as original negatives and positives with contrastive loss. Based on the theoretical analysis about hardness guarantee and limiting behavior, we justify the use of our method. Extensive experiments on retrieval, calibration, few- or zero-shot classification (under distribution shift), embedding arithmetic, and image captioning further show that our method provides transferable representations, enabling robust model adaptation on diverse tasks. Code: https://github.com/changdaeoh/multimodal-mixup

Bumsoo Kim, Jonghwan Mun, Kyoung-Woon On et al.

Human-Object Interaction (HOI) detection is the task of identifying a set of <human, object, interaction> triplets from an image. Recent work proposed transformer encoder-decoder architectures that successfully eliminated the need for many hand-designed components in HOI detection through end-to-end training. However, they are limited to single-scale feature resolution, providing suboptimal performance in scenes containing humans, objects and their interactions with vastly different scales and distances. To tackle this problem, we propose a Multi-Scale TRansformer (MSTR) for HOI detection powered by two novel HOI-aware deformable attention modules called Dual-Entity attention and Entity-conditioned Context attention. While existing deformable attention comes at a huge cost in HOI detection performance, our proposed attention modules of MSTR learn to effectively attend to sampling points that are essential to identify interactions. In experiments, we achieve the new state-of-the-art performance on two HOI detection benchmarks.

Unggi Lee, Minchul Shin, Yeil Jeong et al.

Aligning LLMs for math tutoring typically requires RL-based training with multi-GPU infrastructure. We investigate whether training-free prompt optimization-evolving only the system prompt via API calls-can serve as a practical alternative. We adapt 7 published methods and propose 5 education-specialized methods, evaluating these 12 methods under 5 conditions on 2 OOD benchmark suites. All 12 best-per-method configurations surpass the strongest RL-trained baseline (R_total = 0.633), and our ParetoGrad achieves the best Pareto balance across post-test solve rate, leak control, and helpfulness, rather than dominating any single component. Behavioral analysis with an 82-code educational codebook reveals that training-free methods rely on teaching-knowledge patterns at 2-3x the rate of RL-trained models, with a compensating ~10 percentage-point reduction in intent-level scaffolding. We also find a task-dependent reasoning mode effect consistent across training-free and RL-based paradigms. Our approach enables efficient development of pedagogically aligned LLM tutors with prompts alone and minimal compute.

Jonghwan Mun, Minchul Shin, Gunsoo Han et al.

Self-supervised learning has drawn attention through its effectiveness in learning in-domain representations with no ground-truth annotations; in particular, it is shown that properly designed pretext tasks (e.g., contrastive prediction task) bring significant performance gains for downstream tasks (e.g., classification task). Inspired from this, we tackle video scene segmentation, which is a task of temporally localizing scene boundaries in a video, with a self-supervised learning framework where we mainly focus on designing effective pretext tasks. In our framework, we discover a pseudo-boundary from a sequence of shots by splitting it into two continuous, non-overlapping sub-sequences and leverage the pseudo-boundary to facilitate the pre-training. Based on this, we introduce three novel boundary-aware pretext tasks: 1) Shot-Scene Matching (SSM), 2) Contextual Group Matching (CGM) and 3) Pseudo-boundary Prediction (PP); SSM and CGM guide the model to maximize intra-scene similarity and inter-scene discrimination while PP encourages the model to identify transitional moments. Through comprehensive analysis, we empirically show that pre-training and transferring contextual representation are both critical to improving the video scene segmentation performance. Lastly, we achieve the new state-of-the-art on the MovieNet-SSeg benchmark. The code is available at https://github.com/kakaobrain/bassl.

Geonmo Gu, Byungsoo Ko, SeoungHyun Go et al.

Previous deep learning-based line segment detection (LSD) suffers from the immense model size and high computational cost for line prediction. This constrains them from real-time inference on computationally restricted environments. In this paper, we propose a real-time and light-weight line segment detector for resource-constrained environments named Mobile LSD (M-LSD). We design an extremely efficient LSD architecture by minimizing the backbone network and removing the typical multi-module process for line prediction found in previous methods. To maintain competitive performance with a light-weight network, we present novel training schemes: Segments of Line segment (SoL) augmentation, matching and geometric loss. SoL augmentation splits a line segment into multiple subparts, which are used to provide auxiliary line data during the training process. Moreover, the matching and geometric loss allow a model to capture additional geometric cues. Compared with TP-LSD-Lite, previously the best real-time LSD method, our model (M-LSD-tiny) achieves competitive performance with 2.5% of model size and an increase of 130.5% in inference speed on GPU. Furthermore, our model runs at 56.8 FPS and 48.6 FPS on the latest Android and iPhone mobile devices, respectively. To the best of our knowledge, this is the first real-time deep LSD available on mobile devices. Our code is available.

Minchul Shin, Yoonjae Cho, Byungsoo Ko et al.

In this paper, we study the compositional learning of images and texts for image retrieval. The query is given in the form of an image and text that describes the desired modifications to the image; the goal is to retrieve the target image that satisfies the given modifications and resembles the query by composing information in both the text and image modalities. To remedy this, we propose a novel architecture designed for the image-text composition task and show that the proposed structure can effectively encode the differences between the source and target images conditioned on the text. Furthermore, we introduce a new joint training technique based on the graph convolutional network that is generally applicable for any existing composition methods in a plug-and-play manner. We found that the proposed technique consistently improves performance and achieves state-of-the-art scores on various benchmarks. To avoid misleading experimental results caused by trivial training hyper-parameters, we reproduce all individual baselines and train models with a unified training environment. We expect this approach to suppress undesirable effects from irrelevant components and emphasize the image-text composition module's ability. Also, we achieve the state-of-the-art score without restricting the training environment, which implies the superiority of our method considering the gains from hyper-parameter tuning. The code, including all the baseline methods, are released https://github.com/nashory/rtic-gcn-pytorch.

Minchul Shin, Jonghwan Mun, Kyoung-Woon On et al.

The VALUE (Video-And-Language Understanding Evaluation) benchmark is newly introduced to evaluate and analyze multi-modal representation learning algorithms on three video-and-language tasks: Retrieval, QA, and Captioning. The main objective of the VALUE challenge is to train a task-agnostic model that is simultaneously applicable for various tasks with different characteristics. This technical report describes our winning strategies for the VALUE challenge: 1) single model optimization, 2) transfer learning with visual concepts, and 3) task-aware ensemble. The first and third strategies are designed to address heterogeneous characteristics of each task, and the second one is to leverage rich and fine-grained visual information. We provide a detailed and comprehensive analysis with extensive experimental results. Based on our approach, we ranked first place on the VALUE and QA phases for the competition.

Minchul Shin

This paper presents a study on semi-supervised learning to solve the visual attribute prediction problem. In many applications of vision algorithms, the precise recognition of visual attributes of objects is important but still challenging. This is because defining a class hierarchy of attributes is ambiguous, so training data inevitably suffer from class imbalance and label sparsity, leading to a lack of effective annotations. An intuitive solution is to find a method to effectively learn image representations by utilizing unlabeled images. With that in mind, we propose a multi-teacher-single-student (MTSS) approach inspired by the multi-task learning and the distillation of semi-supervised learning. Our MTSS learns task-specific domain experts called teacher networks using the label embedding technique and learns a unified model called a student network by forcing a model to mimic the distributions learned by domain experts. Our experiments demonstrate that our method not only achieves competitive performance on various benchmarks for fashion attribute prediction, but also improves robustness and cross-domain adaptability for unseen domains.

Minchul Shin, Yoonjae Cho, Seongwuk Hong

This paper is dedicated to team VAA's approach submitted to the Fashion-IQ challenge in CVPR 2020. Given a pair of the image and the text, we present a novel multimodal composition method, RTIC, that can effectively combine the text and the image modalities into a semantic space. We extract the image and the text features that are encoded by the CNNs and the sequential models (e.g., LSTM or GRU), respectively. To emphasize the meaning of the residual of the feature between the target and candidate, the RTIC is composed of N-blocks with channel-wise attention modules. Then, we add the encoded residual to the feature of the candidate image to obtain a synthesized feature. We also explored an ensemble strategy with variants of models and achieved a significant boost in performance comparing to the best single model. Finally, our approach achieved 2nd place in the Fashion-IQ 2020 Challenge with a test score of 48.02 on the leaderboard.

Byungsoo Ko, Minchul Shin, Geonmo Gu et al.

Many studies have been performed on metric learning, which has become a key ingredient in top-performing methods of instance-level image retrieval. Meanwhile, less attention has been paid to pre-processing and post-processing tricks that can significantly boost performance. Furthermore, we found that most previous studies used small scale datasets to simplify processing. Because the behavior of a feature representation in a deep learning model depends on both domain and data, it is important to understand how model behave in large-scale environments when a proper combination of retrieval tricks is used. In this paper, we extensively analyze the effect of well-known pre-processing, post-processing tricks, and their combination for large-scale image retrieval. We found that proper use of these tricks can significantly improve model performance without necessitating complex architecture or introducing loss, as confirmed by achieving a competitive result on the Google Landmark Retrieval Challenge 2019.

Minchul Shin, Sanghyuk Park, Taeksoo Kim

With a growing demand for the search by image, many works have studied the task of fashion instance-level image retrieval (FIR). Furthermore, the recent works introduce a concept of fashion attribute manipulation (FAM) which manipulates a specific attribute (e.g color) of a fashion item while maintaining the rest of the attributes (e.g shape, and pattern). In this way, users can search not only "the same" items but also "similar" items with the desired attributes. FAM is a challenging task in that the attributes are hard to define, and the unique characteristics of a query are hard to be preserved. Although both FIR and FAM are important in real-life applications, most of the previous studies have focused on only one of these problem. In this study, we aim to achieve competitive performance on both FIR and FAM. To do so, we propose a novel method that converts a query into a representation with the desired attributes. We introduce a new idea of attribute manipulation at the feature level, by matching the distribution of manipulated features with real features. In this fashion, the attribute manipulation can be done independently from learning a representation from the image. By introducing the feature-level attribute manipulation, the previous methods for FIR can perform attribute manipulation without sacrificing their retrieval performance.

Minchul Shin, Munsang Kim, Dong-Soo Kwon

We present a baseline convolutional neural network (CNN) structure and image preprocessing methodology to improve facial expression recognition algorithm using CNN. To analyze the most efficient network structure, we investigated four network structures that are known to show good performance in facial expression recognition. Moreover, we also investigated the effect of input image preprocessing methods. Five types of data input (raw, histogram equalization, isotropic smoothing, diffusion-based normalization, difference of Gaussian) were tested, and the accuracy was compared. We trained 20 different CNN models (4 networks x 5 data input types) and verified the performance of each network with test images from five different databases. The experiment result showed that a three-layer structure consisting of a simple convolutional and a max pooling layer with histogram equalization image input was the most efficient. We describe the detailed training procedure and analyze the result of the test accuracy based on considerable observation.