Dahun Kim

Qihang Yu, Huiyu Wang, Dahun Kim et al. · deepmind

We propose Clustering Mask Transformer (CMT-DeepLab), a transformer-based framework for panoptic segmentation designed around clustering. It rethinks the existing transformer architectures used in segmentation and detection; CMT-DeepLab considers the object queries as cluster centers, which fill the role of grouping the pixels when applied to segmentation. The clustering is computed with an alternating procedure, by first assigning pixels to the clusters by their feature affinity, and then updating the cluster centers and pixel features. Together, these operations comprise the Clustering Mask Transformer (CMT) layer, which produces cross-attention that is denser and more consistent with the final segmentation task. CMT-DeepLab improves the performance over prior art significantly by 4.4% PQ, achieving a new state-of-the-art of 55.7% PQ on the COCO test-dev set.

Dahun Kim, Jun Xie, Huiyu Wang et al. · deepmind

We present TubeFormer-DeepLab, the first attempt to tackle multiple core video segmentation tasks in a unified manner. Different video segmentation tasks (e.g., video semantic/instance/panoptic segmentation) are usually considered as distinct problems. State-of-the-art models adopted in the separate communities have diverged, and radically different approaches dominate in each task. By contrast, we make a crucial observation that video segmentation tasks could be generally formulated as the problem of assigning different predicted labels to video tubes (where a tube is obtained by linking segmentation masks along the time axis) and the labels may encode different values depending on the target task. The observation motivates us to develop TubeFormer-DeepLab, a simple and effective video mask transformer model that is widely applicable to multiple video segmentation tasks. TubeFormer-DeepLab directly predicts video tubes with task-specific labels (either pure semantic categories, or both semantic categories and instance identities), which not only significantly simplifies video segmentation models, but also advances state-of-the-art results on multiple video segmentation benchmarks

Weicheng Kuo, AJ Piergiovanni, Dahun Kim et al.

The development of language models have moved from encoder-decoder to decoder-only designs. In addition, we observe that the two most popular multimodal tasks, the generative and contrastive tasks, are nontrivial to accommodate in one architecture, and further need adaptations for downstream tasks. We propose a novel paradigm of training with a decoder-only model for multimodal tasks, which is surprisingly effective in jointly learning of these disparate vision-language tasks. This is done with a simple model, called MaMMUT. It consists of a single vision encoder and a text decoder, and is able to accommodate contrastive and generative learning by a novel two-pass approach on the text decoder. We demonstrate that joint learning of these diverse objectives is simple, effective, and maximizes the weight-sharing of the model across these tasks. Furthermore, the same architecture enables straightforward extensions to open-vocabulary object detection and video-language tasks. The model tackles a diverse range of tasks, while being modest in capacity. Our model achieves the state of the art on image-text and text-image retrieval, video question answering and open-vocabulary detection tasks, outperforming much larger and more extensively trained foundational models. It shows very competitive results on VQA and Video Captioning, especially considering its capacity. Ablations confirm the flexibility and advantages of our approach.

Youngjoong Kwon, Dahun Kim, Duygu Ceylan et al.

We present a method that enables synthesizing novel views and novel poses of arbitrary human performers from sparse multi-view images. A key ingredient of our method is a hybrid appearance blending module that combines the advantages of the implicit body NeRF representation and image-based rendering. Existing generalizable human NeRF methods that are conditioned on the body model have shown robustness against the geometric variation of arbitrary human performers. Yet they often exhibit blurry results when generalized onto unseen identities. Meanwhile, image-based rendering shows high-quality results when sufficient observations are available, whereas it suffers artifacts in sparse-view settings. We propose Neural Image-based Avatars (NIA) that exploits the best of those two methods: to maintain robustness under new articulations and self-occlusions while directly leveraging the available (sparse) source view colors to preserve appearance details of new subject identities. Our hybrid design outperforms recent methods on both in-domain identity generalization as well as challenging cross-dataset generalization settings. Also, in terms of the pose generalization, our method outperforms even the per-subject optimized animatable NeRF methods. The video results are available at https://youngjoongunc.github.io/nia

Runze Li, Dahun Kim, Bir Bhanu et al.

We present RECLIP (Resource-efficient CLIP), a simple method that minimizes computational resource footprint for CLIP (Contrastive Language Image Pretraining). Inspired by the notion of coarse-to-fine in computer vision, we leverage small images to learn from large-scale language supervision efficiently, and finetune the model with high-resolution data in the end. Since the complexity of the vision transformer heavily depends on input image size, our approach significantly reduces the training resource requirements both in theory and in practice. Using the same batch size and training epoch, RECLIP achieves highly competitive zero-shot classification and image-text retrieval accuracy with 6 to 8x less computational resources and 7 to 9x fewer FLOPs than the baseline. Compared to the state-of-the-art contrastive learning methods, RECLIP demonstrates 5 to 59x training resource savings while maintaining highly competitive zero-shot classification and retrieval performance. Finally, RECLIP matches the state of the art in transfer learning to open-vocabulary detection tasks, achieving 32 APr on LVIS. We hope this work will pave the path for the broader research community to explore language supervised pretraining in resource-friendly settings.

AJ Piergiovanni, Isaac Noble, Dahun Kim et al.

One of the main challenges of multimodal learning is the need to combine heterogeneous modalities (e.g., video, audio, text). For example, video and audio are obtained at much higher rates than text and are roughly aligned in time. They are often not synchronized with text, which comes as a global context, e.g., a title, or a description. Furthermore, video and audio inputs are of much larger volumes, and grow as the video length increases, which naturally requires more compute dedicated to these modalities and makes modeling of long-range dependencies harder. We here decouple the multimodal modeling, dividing it into separate, focused autoregressive models, processing the inputs according to the characteristics of the modalities. We propose a multimodal model, called Mirasol3B, consisting of an autoregressive component for the time-synchronized modalities (audio and video), and an autoregressive component for the context modalities which are not necessarily aligned in time but are still sequential. To address the long-sequences of the video-audio inputs, we propose to further partition the video and audio sequences in consecutive snippets and autoregressively process their representations. To that end, we propose a Combiner mechanism, which models the audio-video information jointly within a timeframe. The Combiner learns to extract audio and video features from raw spatio-temporal signals, and then learns to fuse these features producing compact but expressive representations per snippet. Our approach achieves the state-of-the-art on well established multimodal benchmarks, outperforming much larger models. It effectively addresses the high computational demand of media inputs by both learning compact representations, controlling the sequence length of the audio-video feature representations, and modeling their dependencies in time.

Inkyu Shin, Dahun Kim, Qihang Yu et al.

Video Panoptic Segmentation (VPS) aims to achieve comprehensive pixel-level scene understanding by segmenting all pixels and associating objects in a video. Current solutions can be categorized into online and near-online approaches. Evolving over the time, each category has its own specialized designs, making it nontrivial to adapt models between different categories. To alleviate the discrepancy, in this work, we propose a unified approach for online and near-online VPS. The meta architecture of the proposed Video-kMaX consists of two components: within clip segmenter (for clip-level segmentation) and cross-clip associater (for association beyond clips). We propose clip-kMaX (clip k-means mask transformer) and HiLA-MB (Hierarchical Location-Aware Memory Buffer) to instantiate the segmenter and associater, respectively. Our general formulation includes the online scenario as a special case by adopting clip length of one. Without bells and whistles, Video-kMaX sets a new state-of-the-art on KITTI-STEP and VIPSeg for video panoptic segmentation, and VSPW for video semantic segmentation. Code will be made publicly available.

Dahun Kim, Anelia Angelova, Weicheng Kuo

We present a new open-vocabulary detection approach based on region-centric image-language pretraining to bridge the gap between image-level pretraining and open-vocabulary object detection. At the pretraining phase, we incorporate the detector architecture on top of the classification backbone, which better serves the region-level recognition needs of detection by enabling the detector heads to learn from large-scale image-text pairs. Using only standard contrastive loss and no pseudo-labeling, our approach is a simple yet effective extension of the contrastive learning method to learn emergent object-semantic cues. In addition, we propose a shifted-window learning approach upon window attention to make the backbone representation more robust, translation-invariant, and less biased by the window pattern. On the popular LVIS open-vocabulary detection benchmark, our approach sets a new state of the art of 37.6 mask APr using the common ViT-L backbone and public LAION dataset, and 40.5 mask APr using the DataComp-1B dataset, significantly outperforming the best existing approach by +3.7 mask APr at system level. On the COCO benchmark, we achieve very competitive 39.6 novel AP without pseudo labeling or weak supervision. In addition, we evaluate our approach on the transfer detection setup, where it demonstrates notable improvement over the baseline. Visualization reveals emerging object locality from the pretraining recipes compared to the baseline.

Dahun Kim, Ganesh Satish Mallya, Anelia Angelova

Multi-spectral imagery is a valuable input signal for Remote Sensing applications, such as land-use and land-cover classification and environmental monitoring. However, generalist Large Multi-modal Models (LMMs) are typically trained on RGB images, limiting their applicability to the RGB domain. At the same time, training multi-spectral multi-modal models is expensive and produces uniquely specialized models. To address this, we propose a novel training-free approach that introduces multi-spectral data within the inference pipeline of standard RGB-only LMMs, allowing large gains in performance. Our approach leverages the LMMs' understanding of the visual space by adapting non-RGB inputs to that space and injecting domain-specific information and Chain-of-Thought reasoning as instructions. We demonstrate this with the Gemini 2.5 model and observe strong Zero-Shot performance gains on popular Remote Sensing benchmarks. These results highlight the potential for geospatial professionals to leverage powerful generalist models for specialized sensor inputs, benefiting from rich reasoning capabilities grounded in specialized data.

Shijie Wang, Dahun Kim, Ali Taalimi et al.

Visual grounding tasks aim to localize image regions based on natural language references. In this work, we explore whether generative VLMs predominantly trained on image-text data could be leveraged to scale up the text annotation of visual grounding data. We find that grounding knowledge already exists in generative VLM and can be elicited by proper prompting. We thus prompt a VLM to generate object-level descriptions by feeding it object regions from existing object detection datasets. We further propose attribute modeling to explicitly capture the important object attributes, and spatial relation modeling to capture inter-object relationship, both of which are common linguistic pattern in referring expression. Our constructed dataset (500K images, 1M objects, 16M referring expressions) is one of the largest grounding datasets to date, and the first grounding dataset with purely model-generated queries and human-annotated objects. To verify the quality of this data, we conduct zero-shot transfer experiments to the popular RefCOCO benchmarks for both referring expression comprehension (REC) and segmentation (RES) tasks. On both tasks, our model significantly outperform the state-of-the-art approaches without using human annotated visual grounding data. Our results demonstrate the promise of generative VLM to scale up visual grounding in the real world. Code and models will be released.

Minsu Kim, Jeongsoo Choi, Dahun Kim et al.

This paper proposes a textless training method for many-to-many multilingual speech-to-speech translation that can also benefit the transfer of pre-trained knowledge to text-based systems, text-to-speech synthesis and text-to-speech translation. To this end, we represent multilingual speech with speech units that are the discretized representations of speech features derived from a self-supervised speech model. By treating the speech units as pseudo-text, we can focus on the linguistic content of the speech, which can be easily associated with both speech and text modalities at the phonetic level information. By setting both the inputs and outputs of our learning problem as speech units, we propose to train an encoder-decoder model in a many-to-many spoken language translation setting, namely Unit-to-Unit Translation (UTUT). Specifically, the encoder is conditioned on the source language token to correctly understand the input spoken language, while the decoder is conditioned on the target language token to generate the translated speech in the target language. Therefore, during the training, the model can build the knowledge of how languages are comprehended and how to relate them to different languages. Since speech units can be easily associated from both audio and text by quantization and phonemization respectively, the trained model can easily transferred to text-related tasks, even if it is trained in a textless manner. We demonstrate that the proposed UTUT model can be effectively utilized not only for Speech-to-Speech Translation (S2ST) but also for multilingual Text-to-Speech Synthesis (T2S) and Text-to-Speech Translation (T2ST), requiring only minimal fine-tuning steps on text inputs. By conducting comprehensive experiments encompassing various languages, we validate the efficacy of the proposed method across diverse multilingual tasks.

Madhuri Shanbhogue, Zhe Li, Shanfeng Zhang et al.

We introduce Gemini Embedding 2, a native multimodal embedding model that allows embedding video, audio, image, and text modalities in a unified representation space. We leverage the multimodal capabilities of Gemini to produce embeddings for arbitrary combinations of interleaved inputs across all these modalities that generalize well across a wide variety of tasks. Applying large-scale contrastive learning in a multi-task multi-stage training setup, we achieve state-of-the-art performance on key embedding benchmarks including unimodal, cross-modal, and multimodal retrieval spanning a diverse set of tasks. We show that our embedding model demonstrates strong performance (with a score of 62.9 R@1 on MSCOCO, 68.8 NDCG@10 on Vatex, 69.9 on MTEB multilingual and 84.0 on MTEB Code) across a variety of tasks surpassing the performance of specialized models. These unified capabilities make Gemini Embedding 2 a promising candidate for downstream use cases such as RAG, recommendation and search. Furthermore, its robust zero-shot performance across distinct fields - from astronomy and bioscience to fine arts and the culinary arts - establishes it as a highly reliable, out-of-the-box representation even for specialized domains.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Mark Weber, Huiyu Wang, Siyuan Qiao et al.

DeepLab2 is a TensorFlow library for deep labeling, aiming to provide a state-of-the-art and easy-to-use TensorFlow codebase for general dense pixel prediction problems in computer vision. DeepLab2 includes all our recently developed DeepLab model variants with pretrained checkpoints as well as model training and evaluation code, allowing the community to reproduce and further improve upon the state-of-art systems. To showcase the effectiveness of DeepLab2, our Panoptic-DeepLab employing Axial-SWideRNet as network backbone achieves 68.0% PQ or 83.5% mIoU on Cityscaspes validation set, with only single-scale inference and ImageNet-1K pretrained checkpoints. We hope that publicly sharing our library could facilitate future research on dense pixel labeling tasks and envision new applications of this technology. Code is made publicly available at \url{https://github.com/google-research/deeplab2}.

Pierre Musacchio, Jaeyi Jeong, Dahun Kim et al.

Sparse Attention Vectors (SAVs) have emerged as an excellent training-free alternative to supervised finetuning or low-rank adaptation to improve the performance of Vision Language Models (VLMs). At their heart, SAVs select a few accurate attention heads for a task of interest and use them as classifiers, rather than relying on the model's prediction. In a similar spirit, we find that directly probing the raw activations of the VLM, in the form of scalar values, is sufficient to yield accurate classifiers on diverse visually grounded downstream tasks. Shifting focus from attention vectors to scalar activations dramatically increases the search space for accurate parameters, allowing us to find more discriminative neurons immediately from the first generated token. We call such activations Super Neurons (SNs). In this probing setting, we discover that enough SNs appear in the shallower layers of the large language model to allow for extreme early exiting from the first layer of the model at the first generated token. Compared to the original network, SNs robustly improve the classification performance while achieving a speedup of up to 5.10x.

Henrique Schechter Vera, Sahil Dua, Biao Zhang et al.

We introduce EmbeddingGemma, a new lightweight, open text embedding model based on the Gemma 3 language model family. Our innovative training recipe strategically captures knowledge from larger models via encoder-decoder initialization and geometric embedding distillation. We improve model robustness and expressiveness with a spread-out regularizer, and ensure generalizability by merging checkpoints from varied, optimized mixtures. Evaluated on the Massive Text Embedding Benchmark (MTEB) across multilingual, English, and code domains, EmbeddingGemma (300M) achieves state-of-the-art results. Notably, it outperforms prior top models, both proprietary and open, with fewer than 500M parameters, and provides performance comparable to models double its size, offering an exceptional performance-to-cost ratio. Remarkably, this lead persists when quantizing model weights or truncating embedding outputs. This makes EmbeddingGemma particularly well-suited for low-latency and high-throughput use cases such as on-device applications. We provide ablation studies exploring our key design choices. We release EmbeddingGemma to the community to promote further research.

Dahun Kim, AJ Piergiovanni, Ganesh Mallya et al.

We introduce VideoComp, a benchmark and learning framework for advancing video-text compositionality understanding, aimed at improving vision-language models (VLMs) in fine-grained temporal alignment. Unlike existing benchmarks focused on static image-text compositionality or isolated single-event videos, our benchmark targets alignment in continuous multi-event videos. Leveraging video-text datasets with temporally localized event captions (e.g. ActivityNet-Captions, YouCook2), we construct two compositional benchmarks, ActivityNet-Comp and YouCook2-Comp. We create challenging negative samples with subtle temporal disruptions such as reordering, action word replacement, partial captioning, and combined disruptions. These benchmarks comprehensively test models' compositional sensitivity across extended, cohesive video-text sequences. To improve model performance, we propose a hierarchical pairwise preference loss that strengthens alignment with temporally accurate pairs and gradually penalizes increasingly disrupted ones, encouraging fine-grained compositional learning. To mitigate the limited availability of densely annotated video data, we introduce a pretraining strategy that concatenates short video-caption pairs to simulate multi-event sequences. We evaluate video-text foundational models and large multimodal models (LMMs) on our benchmark, identifying both strengths and areas for improvement in compositionality. Overall, our work provides a comprehensive framework for evaluating and enhancing model capabilities in achieving fine-grained, temporally coherent video-text alignment.

AJ Piergiovanni, Dahun Kim, Michael S. Ryoo et al.

Generating automatic dense captions for videos that accurately describe their contents remains a challenging area of research. Most current models require processing the entire video at once. Instead, we propose an efficient, online approach which outputs frequent, detailed and temporally aligned captions, without access to future frames. Our model uses a novel autoregressive factorized decoding architecture, which models the sequence of visual features for each time segment, outputting localized descriptions and efficiently leverages the context from the previous video segments. This allows the model to output frequent, detailed captions to more comprehensively describe the video, according to its actual local content, rather than mimic the training data. Second, we propose an optimization for efficient training and inference, which enables scaling to longer videos. Our approach shows excellent performance compared to both offline and online methods, and uses 20\% less compute. The annotations produced are much more comprehensive and frequent, and can further be utilized in automatic video tagging and in large-scale video data harvesting.

Ganesh Mallya, Yotam Gigi, Dahun Kim et al.

Multi-spectral imagery plays a crucial role in diverse Remote Sensing applications including land-use classification, environmental monitoring and urban planning. These images are widely adopted because their additional spectral bands correlate strongly with physical materials on the ground, such as ice, water, and vegetation. This allows for more accurate identification, and their public availability from missions, such as Sentinel-2 and Landsat, only adds to their value. Currently, the automatic analysis of such data is predominantly managed through machine learning models specifically trained for multi-spectral input, which are costly to train and support. Furthermore, although providing a lot of utility for Remote Sensing, such additional inputs cannot be used with powerful generalist large multimodal models, which are capable of solving many visual problems, but are not able to understand specialized multi-spectral signals. To address this, we propose a training-free approach which introduces new multi-spectral data in a Zero-Shot-only mode, as inputs to generalist multimodal models, trained on RGB-only inputs. Our approach leverages the multimodal models' understanding of the visual space, and proposes to adapt to inputs to that space, and to inject domain-specific information as instructions into the model. We exemplify this idea with the Gemini2.5 model and observe strong Zero-Shot performance gains of the approach on popular Remote Sensing benchmarks for land cover and land use classification and demonstrate the easy adaptability of Gemini2.5 to new inputs. These results highlight the potential for geospatial professionals, working with non-standard specialized inputs, to easily leverage powerful multimodal models, such as Gemini2.5, to accelerate their work, benefiting from their rich reasoning and contextual capabilities, grounded in the specialized sensor data.

AJ Piergiovanni, Ganesh Satish Mallya, Dahun Kim et al.

Dense video captioning is a challenging video understanding task which aims to simultaneously segment the video into a sequence of meaningful consecutive events and to generate detailed captions to accurately describe each event. Existing methods often encounter difficulties when working with the long videos associated with dense video captioning, due to the computational complexity and memory limitations. Furthermore, traditional approaches require the entire video as input, in order to produce an answer, which precludes online processing of the video. We address these challenges by time-scaling State-Space Models (SSMs) to even longer sequences than before. Our approach, State-Space Models with Transfer State, combines both the long-sequence and recurrent properties of SSMs and addresses the main limitation of SSMs which are otherwise not able to sustain their state for very long contexts, effectively scaling SSMs further in time. The proposed model is particularly suitable for generating captions on-the-fly, in an online or streaming manner, without having to wait for the full video to be processed, which is more beneficial in practice. When applied to dense video captioning, our approach scales well with video lengths and uses 7x fewer FLOPs.

Dahun Kim, Anelia Angelova

We propose Context-Adaptive Multi-Prompt Embedding, a novel approach to enrich semantic representations in vision-language contrastive learning. Unlike standard CLIP-style models that rely on a single text embedding, our method introduces multiple structured prompts, each containing a distinct adaptive token that captures diverse semantic aspects of the input text. We leverage a pretrained LLM as the text encoder within the CLIP framework, processing all prompts jointly in a single forward pass. The resulting prompt embeddings are combined into a unified text representation, enabling semantically richer alignment with visual features. To further promote semantic diversity and representation quality, we incorporate a diversity regularization loss and a negation-aware loss, encouraging specialization across prompts and improving contrastive discrimination. Our method achieves consistent improvements on both image-text and video-text retrieval benchmarks.

Dahun Kim, Anelia Angelova, Weicheng Kuo

We present Contrastive Feature Masking Vision Transformer (CFM-ViT) - an image-text pretraining methodology that achieves simultaneous learning of image- and region-level representation for open-vocabulary object detection (OVD). Our approach combines the masked autoencoder (MAE) objective into the contrastive learning objective to improve the representation for localization tasks. Unlike standard MAE, we perform reconstruction in the joint image-text embedding space, rather than the pixel space as is customary with the classical MAE method, which causes the model to better learn region-level semantics. Moreover, we introduce Positional Embedding Dropout (PED) to address scale variation between image-text pretraining and detection finetuning by randomly dropping out the positional embeddings during pretraining. PED improves detection performance and enables the use of a frozen ViT backbone as a region classifier, preventing the forgetting of open-vocabulary knowledge during detection finetuning. On LVIS open-vocabulary detection benchmark, CFM-ViT achieves a state-of-the-art 33.9 AP$r$, surpassing the best approach by 7.6 points and achieves better zero-shot detection transfer. Finally, CFM-ViT acquires strong image-level representation, outperforming the state of the art on 8 out of 12 metrics on zero-shot image-text retrieval benchmarks.

Dahun Kim, Anelia Angelova, Weicheng Kuo

We present Region-aware Open-vocabulary Vision Transformers (RO-ViT) - a contrastive image-text pretraining recipe to bridge the gap between image-level pretraining and open-vocabulary object detection. At the pretraining phase, we propose to randomly crop and resize regions of positional embeddings instead of using the whole image positional embeddings. This better matches the use of positional embeddings at region-level in the detection finetuning phase. In addition, we replace the common softmax cross entropy loss in contrastive learning with focal loss to better learn the informative yet difficult examples. Finally, we leverage recent advances in novel object proposals to improve open-vocabulary detection finetuning. We evaluate our full model on the LVIS and COCO open-vocabulary detection benchmarks and zero-shot transfer. RO-ViT achieves a state-of-the-art 34.1 $AP_r$ on LVIS, surpassing the best existing approach by +7.8 points in addition to competitive zero-shot transfer detection. Surprisingly, RO-ViT improves the image-level representation as well and achieves the state of the art on 9 out of 12 metrics on COCO and Flickr image-text retrieval benchmarks, outperforming competitive approaches with larger models.

Youngjoong Kwon, Dahun Kim, Duygu Ceylan et al.

In this paper, we aim at synthesizing a free-viewpoint video of an arbitrary human performance using sparse multi-view cameras. Recently, several works have addressed this problem by learning person-specific neural radiance fields (NeRF) to capture the appearance of a particular human. In parallel, some work proposed to use pixel-aligned features to generalize radiance fields to arbitrary new scenes and objects. Adopting such generalization approaches to humans, however, is highly challenging due to the heavy occlusions and dynamic articulations of body parts. To tackle this, we propose Neural Human Performer, a novel approach that learns generalizable neural radiance fields based on a parametric human body model for robust performance capture. Specifically, we first introduce a temporal transformer that aggregates tracked visual features based on the skeletal body motion over time. Moreover, a multi-view transformer is proposed to perform cross-attention between the temporally-fused features and the pixel-aligned features at each time step to integrate observations on the fly from multiple views. Experiments on the ZJU-MoCap and AIST datasets show that our method significantly outperforms recent generalizable NeRF methods on unseen identities and poses. The video results and code are available at https://youngjoongunc.github.io/nhp.

Dahun Kim, Tsung-Yi Lin, Anelia Angelova et al.

Object proposals have become an integral preprocessing steps of many vision pipelines including object detection, weakly supervised detection, object discovery, tracking, etc. Compared to the learning-free methods, learning-based proposals have become popular recently due to the growing interest in object detection. The common paradigm is to learn object proposals from data labeled with a set of object regions and their corresponding categories. However, this approach often struggles with novel objects in the open world that are absent in the training set. In this paper, we identify that the problem is that the binary classifiers in existing proposal methods tend to overfit to the training categories. Therefore, we propose a classification-free Object Localization Network (OLN) which estimates the objectness of each region purely by how well the location and shape of a region overlap with any ground-truth object (e.g., centerness and IoU). This simple strategy learns generalizable objectness and outperforms existing proposals on cross-category generalization on COCO, as well as cross-dataset evaluation on RoboNet, Object365, and EpicKitchens. Finally, we demonstrate the merit of OLN for long-tail object detection on large vocabulary dataset, LVIS, where we notice clear improvement in rare and common categories.

Sanghyun Woo, Dahun Kim, Joon-Young Lee et al.

Temporal correspondence - linking pixels or objects across frames - is a fundamental supervisory signal for the video models. For the panoptic understanding of dynamic scenes, we further extend this concept to every segment. Specifically, we aim to learn coarse segment-level matching and fine pixel-level matching together. We implement this idea by designing two novel learning objectives. To validate our proposals, we adopt a deep siamese model and train the model to learn the temporal correspondence on two different levels (i.e., segment and pixel) along with the target task. At inference time, the model processes each frame independently without any extra computation and post-processing. We show that our per-frame inference model can achieve new state-of-the-art results on Cityscapes-VPS and VIPER datasets. Moreover, due to its high efficiency, the model runs in a fraction of time (3x) compared to the previous state-of-the-art approach.

Myungchul Kim, Sanghyun Woo, Dahun Kim et al.

Pursuing a more coherent scene understanding towards real-time vision applications, single-stage instance segmentation has recently gained popularity, achieving a simpler and more efficient design than its two-stage counterparts. Besides, its global mask representation often leads to superior accuracy to the two-stage Mask R-CNN which has been dominant thus far. Despite the promising advances in single-stage methods, finer delineation of instance boundaries still remains unexcavated. Indeed, boundary information provides a strong shape representation that can operate in synergy with the fully-convolutional mask features of the single-stage segmenter. In this work, we propose Boundary Basis based Instance Segmentation(B2Inst) to learn a global boundary representation that can complement existing global-mask-based methods that are often lacking high-frequency details. Besides, we devise a unified quality measure of both mask and boundary and introduce a network block that learns to score the per-instance predictions of itself. When applied to the strongest baselines in single-stage instance segmentation, our B2Inst leads to consistent improvements and accurately parse out the instance boundaries in a scene. Regardless of being single-stage or two-stage frameworks, we outperform the existing state-of-the-art methods on the COCO dataset with the same ResNet-50 and ResNet-101 backbones.

Dahun Kim, Sanghyun Woo, Joon-Young Lee et al.

Panoptic segmentation has become a new standard of visual recognition task by unifying previous semantic segmentation and instance segmentation tasks in concert. In this paper, we propose and explore a new video extension of this task, called video panoptic segmentation. The task requires generating consistent panoptic segmentation as well as an association of instance ids across video frames. To invigorate research on this new task, we present two types of video panoptic datasets. The first is a re-organization of the synthetic VIPER dataset into the video panoptic format to exploit its large-scale pixel annotations. The second is a temporal extension on the Cityscapes val. set, by providing new video panoptic annotations (Cityscapes-VPS). Moreover, we propose a novel video panoptic segmentation network (VPSNet) which jointly predicts object classes, bounding boxes, masks, instance id tracking, and semantic segmentation in video frames. To provide appropriate metrics for this task, we propose a video panoptic quality (VPQ) metric and evaluate our method and several other baselines. Experimental results demonstrate the effectiveness of the presented two datasets. We achieve state-of-the-art results in image PQ on Cityscapes and also in VPQ on Cityscapes-VPS and VIPER datasets. The datasets and code are made publicly available.

Yunjae Jung, Dahun Kim, Sanghyun Woo et al.

Visual storytelling is a task of creating a short story based on photo streams. Unlike existing visual captioning, storytelling aims to contain not only factual descriptions, but also human-like narration and semantics. However, the VIST dataset consists only of a small, fixed number of photos per story. Therefore, the main challenge of visual storytelling is to fill in the visual gap between photos with narrative and imaginative story. In this paper, we propose to explicitly learn to imagine a storyline that bridges the visual gap. During training, one or more photos is randomly omitted from the input stack, and we train the network to produce a full plausible story even with missing photo(s). Furthermore, we propose for visual storytelling a hide-and-tell model, which is designed to learn non-local relations across the photo streams and to refine and improve conventional RNN-based models. In experiments, we show that our scheme of hide-and-tell, and the network design are indeed effective at storytelling, and that our model outperforms previous state-of-the-art methods in automatic metrics. Finally, we qualitatively show the learned ability to interpolate storyline over visual gaps.

Kwanyong Park, Sanghyun Woo, Dahun Kim et al.

In this paper, we investigate the problem of unpaired video-to-video translation. Given a video in the source domain, we aim to learn the conditional distribution of the corresponding video in the target domain, without seeing any pairs of corresponding videos. While significant progress has been made in the unpaired translation of images, directly applying these methods to an input video leads to low visual quality due to the additional time dimension. In particular, previous methods suffer from semantic inconsistency (i.e., semantic label flipping) and temporal flickering artifacts. To alleviate these issues, we propose a new framework that is composed of carefully-designed generators and discriminators, coupled with two core objective functions: 1) content preserving loss and 2) temporal consistency loss. Extensive qualitative and quantitative evaluations demonstrate the superior performance of the proposed method against previous approaches. We further apply our framework to a domain adaptation task and achieve favorable results.

Sanghyun Woo, Dahun Kim, KwanYong Park et al.

We propose a novel feed-forward network for video inpainting. We use a set of sampled video frames as the reference to take visible contents to fill the hole of a target frame. Our video inpainting network consists of two stages. The first stage is an alignment module that uses computed homographies between the reference frames and the target frame. The visible patches are then aggregated based on the frame similarity to fill in the target holes roughly. The second stage is a non-local attention module that matches the generated patches with known reference patches (in space and time) to refine the previous global alignment stage. Both stages consist of large spatial-temporal window size for the reference and thus enable modeling long-range correlations between distant information and the hole regions. Therefore, even challenging scenes with large or slowly moving holes can be handled, which have been hardly modeled by existing flow-based approach. Our network is also designed with a recurrent propagation stream to encourage temporal consistency in video results. Experiments on video object removal demonstrate that our method inpaints the holes with globally and locally coherent contents.

Dahun Kim, Sanghyun Woo, Joon-Young Lee et al.

Blind video decaptioning is a problem of automatically removing text overlays and inpainting the occluded parts in videos without any input masks. While recent deep learning based inpainting methods deal with a single image and mostly assume that the positions of the corrupted pixels are known, we aim at automatic text removal in video sequences without mask information. In this paper, we propose a simple yet effective framework for fast blind video decaptioning. We construct an encoder-decoder model, where the encoder takes multiple source frames that can provide visible pixels revealed from the scene dynamics. These hints are aggregated and fed into the decoder. We apply a residual connection from the input frame to the decoder output to enforce our network to focus on the corrupted regions only. Our proposed model was ranked in the first place in the ECCV Chalearn 2018 LAP Inpainting Competition Track2: Video decaptioning. In addition, we further improve this strong model by applying a recurrent feedback. The recurrent feedback not only enforces temporal coherence but also provides strong clues on where the corrupted pixels are. Both qualitative and quantitative experiments demonstrate that our full model produces accurate and temporally consistent video results in real time (50+ fps).

Dahun Kim, Sanghyun Woo, Joon-Young Lee et al.

Video inpainting aims to fill spatio-temporal holes with plausible content in a video. Despite tremendous progress of deep neural networks for image inpainting, it is challenging to extend these methods to the video domain due to the additional time dimension. In this work, we propose a novel deep network architecture for fast video inpainting. Built upon an image-based encoder-decoder model, our framework is designed to collect and refine information from neighbor frames and synthesize still-unknown regions. At the same time, the output is enforced to be temporally consistent by a recurrent feedback and a temporal memory module. Compared with the state-of-the-art image inpainting algorithm, our method produces videos that are much more semantically correct and temporally smooth. In contrast to the prior video completion method which relies on time-consuming optimization, our method runs in near real-time while generating competitive video results. Finally, we applied our framework to video retargeting task, and obtain visually pleasing results.

Dahun Kim, Donghyeon Cho, In So Kweon

Self-supervised tasks such as colorization, inpainting and zigsaw puzzle have been utilized for visual representation learning for still images, when the number of labeled images is limited or absent at all. Recently, this worthwhile stream of study extends to video domain where the cost of human labeling is even more expensive. However, the most of existing methods are still based on 2D CNN architectures that can not directly capture spatio-temporal information for video applications. In this paper, we introduce a new self-supervised task called as \textit{Space-Time Cubic Puzzles} to train 3D CNNs using large scale video dataset. This task requires a network to arrange permuted 3D spatio-temporal crops. By completing \textit{Space-Time Cubic Puzzles}, the network learns both spatial appearance and temporal relation of video frames, which is our final goal. In experiments, we demonstrate that our learned 3D representation is well transferred to action recognition tasks, and outperforms state-of-the-art 2D CNN-based competitors on UCF101 and HMDB51 datasets.

Yunjae Jung, Donghyeon Cho, Dahun Kim et al.

In this paper, we address the problem of unsupervised video summarization that automatically extracts key-shots from an input video. Specifically, we tackle two critical issues based on our empirical observations: (i) Ineffective feature learning due to flat distributions of output importance scores for each frame, and (ii) training difficulty when dealing with long-length video inputs. To alleviate the first problem, we propose a simple yet effective regularization loss term called variance loss. The proposed variance loss allows a network to predict output scores for each frame with high discrepancy which enables effective feature learning and significantly improves model performance. For the second problem, we design a novel two-stream network named Chunk and Stride Network (CSNet) that utilizes local (chunk) and global (stride) temporal view on the video features. Our CSNet gives better summarization results for long-length videos compared to the existing methods. In addition, we introduce an attention mechanism to handle the dynamic information in videos. We demonstrate the effectiveness of the proposed methods by conducting extensive ablation studies and show that our final model achieves new state-of-the-art results on two benchmark datasets.

Sanghyun Woo, Dahun Kim, Donghyeon Cho et al.

Objects and their relationships are critical contents for image understanding. A scene graph provides a structured description that captures these properties of an image. However, reasoning about the relationships between objects is very challenging and only a few recent works have attempted to solve the problem of generating a scene graph from an image. In this paper, we present a method that improves scene graph generation by explicitly modeling inter-dependency among the entire object instances. We design a simple and effective relational embedding module that enables our model to jointly represent connections among all related objects, rather than focus on an object in isolation. Our method significantly benefits the main part of the scene graph generation task: relationship classification. Using it on top of a basic Faster R-CNN, our model achieves state-of-the-art results on the Visual Genome benchmark. We further push the performance by introducing global context encoding module and geometrical layout encoding module. We validate our final model, LinkNet, through extensive ablation studies, demonstrating its efficacy in scene graph generation.

Dahun Kim, Donghyeon Cho, Donggeun Yoo et al.

In this paper, we explore methods of complicating self-supervised tasks for representation learning. That is, we do severe damage to data and encourage a network to recover them. First, we complicate each of three powerful self-supervised task candidates: jigsaw puzzle, inpainting, and colorization. In addition, we introduce a novel complicated self-supervised task called "Completing damaged jigsaw puzzles" which is puzzles with one piece missing and the other pieces without color. We train a convolutional neural network not only to solve the puzzles, but also generate the missing content and colorize the puzzles. The recovery of the aforementioned damage pushes the network to obtain robust and general-purpose representations. We demonstrate that complicating the self-supervised tasks improves their original versions and that our final task learns more robust and transferable representations compared to the previous methods, as well as the simple combination of our candidate tasks. Our approach achieves state-of-the-art performance in transfer learning on PASCAL classification and semantic segmentation.

Dahun Kim, Donghyeon Cho, Donggeun Yoo et al.

Weakly supervised semantic segmentation and localiza- tion have a problem of focusing only on the most important parts of an image since they use only image-level annota- tions. In this paper, we solve this problem fundamentally via two-phase learning. Our networks are trained in two steps. In the first step, a conventional fully convolutional network (FCN) is trained to find the most discriminative parts of an image. In the second step, the activations on the most salient parts are suppressed by inference conditional feedback, and then the second learning is performed to find the area of the next most important parts. By combining the activations of both phases, the entire portion of the tar- get object can be captured. Our proposed training scheme is novel and can be utilized in well-designed techniques for weakly supervised semantic segmentation, salient region detection, and object location prediction. Detailed experi- ments demonstrate the effectiveness of our two-phase learn- ing in each task.