NSTL国家科技图书文献中心

1. ShareGPT4V: Improving Large Multi-modal Models with Better Captions NSTL国家科技图书文献中心

Lin Chen | Jinsong Li ... - 《Computer Vision - ECCV 2024,Part XVII》 - European Conference on Computer Vision - 2025, - 370～387 - 共18页

摘要： detailed captions* enable more nuanced vision-language* | brief captions* or VQA data; 2) Cutting-edge LMMs can* | captions, exhibiting remarkable improvements across a | range from thousands to millions of captions. Drawing | ShareGPT4V, consisting of 100K high-quality captions

关键词： Large multi-modal models | Modality alignment | High-quality captions

2. Parrot Captions Teach CLIP to Spot Text NSTL国家科技图书文献中心

Yiqi Lin | Conghui He ... - 《Computer Vision - ECCV 2024,Part XLII》 - European Conference on Computer Vision - 2025, - 368～385 - 共18页

摘要： dataset LAION-2B, the captions* also densely parrot* | text content and around 30% of captions* words are* | whether these parrot captions* shape the text spotting* | criteria. We show that training with parrot captions | Despite CLIP being the foundation model in

关键词： Image-Text dataset | Text spotting bias

3. DreamLIP: Language-Image Pre-training with Long Captions NSTL国家科技图书文献中心

Kecheng Zheng | Yifei Zhang ... - 《Computer Vision - ECCV 2024,Part XVIII》 - European Conference on Computer Vision - 2025, - 73～90 - 共18页

摘要： requires lengthy captions* (e.g., with 10 sentences* | benefit from long captions. To figure this out, we first | resulting captions* under a contrastive learning framework* | dynamically sample sub-captions* from the text label to* | Language-image pre-training largely relies on

关键词： Language-image pre-training | Long caption | Multi-modal learning

4. Misinformation in Reels, Influence of Contextual Superimposed Texts in Short Videos NSTL国家科技图书文献中心

Andrew Bartlett | Waheeb Yaqub ... - 《Web Information Systems Engineering - WISE 2024,Part II》 - International Conference on Web Information Systems Engineering - 2025, - 3～14 - 共12页

摘要： captions that create doubt in viewers' minds. With the | Australia to examine the impact of misleading captions* on* | Short videos and reels have emerged as a new | method for conveying visual information on social media | . These videos often include misleading superimposed

关键词： Fake short videos | Fake news | Social media | Context captions | Perceptions | Misinformation

5. MCANet: Multimodal Caption Aware Training-Free Video Anomaly Detection via Large Language Model NSTL国家科技图书文献中心

Prabhu Prasad Dev | Raju Hazari ... - 《Pattern Recognition,Part XXXII》 - International Conference on Pattern Recognition - 2025, - 362～379 - 共18页

摘要： captions produced by the audio captioning model. The | utilizes image-text similarities to clean noisy captions | generated by the image captioning* model, while the second* | Towards Video Anomaly Detection (VAD | ), existing methods require labor-intensive data collection

关键词： Video anomaly detection | Large language model | Vision language model | Audio language model | Multimodal captions

6. Interactive Video Search with Multi-modal LLM Video Captioning NSTL国家科技图书文献中心

Yu-Tong Cheng | Jiaxin Wu ... - 《MultiMedia Modeling,Part V》 - International Conference on MultiMedia Modeling - 2025, - 302～309 - 共8页

摘要： use LLM to generate video captions* for a large video* | -grained captions* for test video collections to enable* | detailed captions* in our interactive video retrieval* | captions are effective in improving the search accuracy | multi-model LLM on video captioning. Specifically, we

关键词： Interactive video retrieval | Multi-modal LLM | Video captioning

7. Controllable Contextualized Image Captioning: Directing the Visual Narrative Through User-Defined Highlights NSTL国家科技图书文献中心

Shunqi Mao | Chaoyi Zhang ... - 《Computer Vision - ECCV 2024,Part L》 - European Conference on Computer Vision - 2025, - 464～481 - 共18页

摘要： generate image captions* given specific contextual* | the model to tailor captions* that resonate with the* | captions. P-Ctrl conditions the model generation on | highlight by prepending captions* with highlight-driven* | assess the quality of the controlled captions* alongside*

关键词： Contextualized image captioning | Large multimodal model | Controllable text generation

8. It's Just Another Day: Unique Video Captioning by Discriminitive Prompting NSTL国家科技图书文献中心

Toby Perrett | Tengda Han ... - 《Computer Vision - ACCV 2024,Part III》 - Asian Conference on Computer Vision - 2025, - 275～293 - 共19页

摘要： given identical captions, which makes it difficult to | generate unique captions. We introduce two benchmarks for | demonstrate that captions* generated by CDP improve text-to* | captioning: Given multiple clips with the same caption, we | identifies it. We propose Captioning* by Discriminative*

关键词： Uniqueness | Video captioning | Egocentric | Movies

9. Deneb: A Hallucination-Robust Automatic Evaluation Metric for Image Captioning NSTL国家科技图书文献中心

Kazuki Matsuda | Yuiga Wada ... - 《Computer Vision - ACCV 2024,Part III》 - Asian Conference on Computer Vision - 2025, - 166～182 - 共17页

摘要： ability to compare candidate captions* with multifaceted* | reference captions. To address this shortcoming, we | captions. To train Deneb, we construct the diverse and | captioning, with a particular focus on robustness against | In this work, we address the challenge of

关键词： Vision and language | Hallucination | Image captioning | Metrics

10. ControlCap: Controllable Region-Level Captioning NSTL国家科技图书文献中心

Yuzhong Zhao | Yue Liu ... - 《Computer Vision - ECCV 2024,Part XXXVIII》 - European Conference on Computer Vision - 2025, - 21～38 - 共18页

摘要： frequent captions* but miss the less frequent ones. In* | captions within a few sub-spaces containing the control | frequent captions, alleviating the caption degeneration | Region-level captioning* is challenged by the* | captioning (Control-Cap) approach, which introduces control

摘要： detailed captions* enable more nuanced vision-language* | brief captions* or VQA data; 2) Cutting-edge LMMs can* | captions, exhibiting remarkable improvements across a | range from thousands to millions of captions. Drawing | ShareGPT4V, consisting of 100K high-quality captions

关键词： Large multi-modal models | Modality alignment | High-quality captions

摘要： dataset LAION-2B, the captions* also densely parrot* | text content and around 30% of captions* words are* | whether these parrot captions* shape the text spotting* | criteria. We show that training with parrot captions | Despite CLIP being the foundation model in

关键词： Image-Text dataset | Text spotting bias

摘要： requires lengthy captions* (e.g., with 10 sentences* | benefit from long captions. To figure this out, we first | resulting captions* under a contrastive learning framework* | dynamically sample sub-captions* from the text label to* | Language-image pre-training largely relies on

关键词： Language-image pre-training | Long caption | Multi-modal learning

摘要： captions that create doubt in viewers' minds. With the | Australia to examine the impact of misleading captions* on* | Short videos and reels have emerged as a new | method for conveying visual information on social media | . These videos often include misleading superimposed

关键词： Fake short videos | Fake news | Social media | Context captions | Perceptions | Misinformation

摘要： captions produced by the audio captioning model. The | utilizes image-text similarities to clean noisy captions | generated by the image captioning* model, while the second* | Towards Video Anomaly Detection (VAD | ), existing methods require labor-intensive data collection

关键词： Video anomaly detection | Large language model | Vision language model | Audio language model | Multimodal captions

摘要： use LLM to generate video captions* for a large video* | -grained captions* for test video collections to enable* | detailed captions* in our interactive video retrieval* | captions are effective in improving the search accuracy | multi-model LLM on video captioning. Specifically, we

关键词： Interactive video retrieval | Multi-modal LLM | Video captioning

摘要： generate image captions* given specific contextual* | the model to tailor captions* that resonate with the* | captions. P-Ctrl conditions the model generation on | highlight by prepending captions* with highlight-driven* | assess the quality of the controlled captions* alongside*

关键词： Contextualized image captioning | Large multimodal model | Controllable text generation

关键词： Uniqueness | Video captioning | Egocentric | Movies

摘要： ability to compare candidate captions* with multifaceted* | reference captions. To address this shortcoming, we | captions. To train Deneb, we construct the diverse and | captioning, with a particular focus on robustness against | In this work, we address the challenge of

关键词： Vision and language | Hallucination | Image captioning | Metrics

摘要： frequent captions* but miss the less frequent ones. In* | captions within a few sub-spaces containing the control | frequent captions, alleviating the caption degeneration | Region-level captioning* is challenged by the* | captioning (Control-Cap) approach, which introduces control

关键词： Controllable captioning | Caption degeneration | Region-level captioning

4008-161-200 800-990-8900

国家科技图书文献中心

摘要： detailed captions enable more nuanced vision-language | brief captions or VQA data; 2) Cutting-edge LMMs can | captions, exhibiting remarkable improvements across a | range from thousands to millions of captions. Drawing | ShareGPT4V, consisting of 100K high-quality captions

关键词： Large multi-modal models | Modality alignment | High-quality captions

摘要： dataset LAION-2B, the captions also densely parrot | text content and around 30% of captions words are | whether these parrot captions shape the text spotting | criteria. We show that training with parrot captions | Despite CLIP being the foundation model in

关键词： Image-Text dataset | Text spotting bias

摘要： requires lengthy captions (e.g., with 10 sentences | benefit from long captions. To figure this out, we first | resulting captions under a contrastive learning framework | dynamically sample sub-captions from the text label to | Language-image pre-training largely relies on

关键词： Language-image pre-training | Long caption | Multi-modal learning

摘要： captions that create doubt in viewers' minds. With the | Australia to examine the impact of misleading captions on | Short videos and reels have emerged as a new | method for conveying visual information on social media | . These videos often include misleading superimposed

关键词： Fake short videos | Fake news | Social media | Context captions | Perceptions | Misinformation

摘要： captions produced by the audio captioning model. The | utilizes image-text similarities to clean noisy captions | generated by the image captioning model, while the second | Towards Video Anomaly Detection (VAD | ), existing methods require labor-intensive data collection

关键词： Video anomaly detection | Large language model | Vision language model | Audio language model | Multimodal captions

摘要： use LLM to generate video captions for a large video | -grained captions for test video collections to enable | detailed captions in our interactive video retrieval | captions are effective in improving the search accuracy | multi-model LLM on video captioning. Specifically, we

关键词： Interactive video retrieval | Multi-modal LLM | Video captioning

摘要： generate image captions given specific contextual | the model to tailor captions that resonate with the | captions. P-Ctrl conditions the model generation on | highlight by prepending captions with highlight-driven | assess the quality of the controlled captions alongside

关键词： Contextualized image captioning | Large multimodal model | Controllable text generation

关键词： Uniqueness | Video captioning | Egocentric | Movies

摘要： ability to compare candidate captions with multifaceted | reference captions. To address this shortcoming, we | captions. To train Deneb, we construct the diverse and | captioning, with a particular focus on robustness against | In this work, we address the challenge of

关键词： Vision and language | Hallucination | Image captioning | Metrics

摘要： frequent captions but miss the less frequent ones. In | captions within a few sub-spaces containing the control | frequent captions, alleviating the caption degeneration | Region-level captioning is challenged by the | captioning (Control-Cap) approach, which introduces control

关键词： Controllable captioning | Caption degeneration | Region-level captioning

4008-161-200 800-990-8900

国家科技图书文献中心

摘要： detailed captions* enable more nuanced vision-language* | brief captions* or VQA data; 2) Cutting-edge LMMs can* | captions, exhibiting remarkable improvements across a | range from thousands to millions of captions. Drawing | ShareGPT4V, consisting of 100K high-quality captions

摘要： dataset LAION-2B, the captions* also densely parrot* | text content and around 30% of captions* words are* | whether these parrot captions* shape the text spotting* | criteria. We show that training with parrot captions | Despite CLIP being the foundation model in

摘要： requires lengthy captions* (e.g., with 10 sentences* | benefit from long captions. To figure this out, we first | resulting captions* under a contrastive learning framework* | dynamically sample sub-captions* from the text label to* | Language-image pre-training largely relies on

摘要： captions that create doubt in viewers' minds. With the | Australia to examine the impact of misleading captions* on* | Short videos and reels have emerged as a new | method for conveying visual information on social media | . These videos often include misleading superimposed

摘要： captions produced by the audio captioning model. The | utilizes image-text similarities to clean noisy captions | generated by the image captioning* model, while the second* | Towards Video Anomaly Detection (VAD | ), existing methods require labor-intensive data collection

摘要： use LLM to generate video captions* for a large video* | -grained captions* for test video collections to enable* | detailed captions* in our interactive video retrieval* | captions are effective in improving the search accuracy | multi-model LLM on video captioning. Specifically, we

摘要： generate image captions* given specific contextual* | the model to tailor captions* that resonate with the* | captions. P-Ctrl conditions the model generation on | highlight by prepending captions* with highlight-driven* | assess the quality of the controlled captions* alongside*

摘要： ability to compare candidate captions* with multifaceted* | reference captions. To address this shortcoming, we | captions. To train Deneb, we construct the diverse and | captioning, with a particular focus on robustness against | In this work, we address the challenge of

摘要： frequent captions* but miss the less frequent ones. In* | captions within a few sub-spaces containing the control | frequent captions, alleviating the caption degeneration | Region-level captioning* is challenged by the* | captioning (Control-Cap) approach, which introduces control