A NeRF-Based Captioning Framework for Spatially Rich and Context-Aware Image Descriptions

Đã lưu trong:
Chi tiết về thư mục
Xuất bản năm:Journal Europeen des Systemes Automatises vol. 58, no. 5 (May 2025), p. 1059-1065
Tác giả chính: Garine, Bindu Madhavi
Tác giả khác: Parimala, Rajeshwari, Motukuri, Sridevi, Raja Sekhar Reddy Pocha, Pulipati, Srilatha
Được phát hành:
International Information and Engineering Technology Association (IIETA)
Những chủ đề:
Truy cập trực tuyến:Citation/Abstract
Full Text - PDF
Các nhãn: Thêm thẻ
Không có thẻ, Là người đầu tiên thẻ bản ghi này!

MARC

LEADER 00000nab a2200000uu 4500
001 3231508339
003 UK-CbPIL
022 |a 1269-6935 
022 |a 2116-7087 
024 7 |a 10.18280/jesa.580518  |2 doi 
035 |a 3231508339 
045 2 |b d20250501  |b d20250531 
100 1 |a Garine, Bindu Madhavi 
245 1 |a A NeRF-Based Captioning Framework for Spatially Rich and Context-Aware Image Descriptions 
260 |b International Information and Engineering Technology Association (IIETA)  |c May 2025 
513 |a Journal Article 
520 3 |a Traditional caption models are mainly dependent on 2D visual properties, which limit their ability to understand and describe spatial conditions, depth and three-dimensional structures in images. These models struggle to capture object interviews, beef and light variations, which are important for generating relevant and spatial conscious details. To address these boundaries, we introduce Neural Radiance Feilds Captioning (NeRF-Cap) framework is a new Neural Radiance Field based on multimodal image-tight frame that integrates 3D-visual reconstruction with natural language treatment (NLP). NeRF's ability to create a constant volumetric representation of a view of several 2D approaches enables the recovery of depth-individual and geometrically accurate functions, which improves the descriptive power of the caption generated. Our approach also integrates the advanced visual language models such as Bootstrapping Language-Image Pre-training (BLIP), Contrastive Language-Image Pretraining (CLIP) and Large Language Model Meta AI (LLaMA) which process the text details by involving semantic object interlation, depth such and light effect in the caption process. By taking advantage of the high definition 3D representation of the NeRF, NeRF-Cap improved traditional captions by providing spatial consistent, photorealist and geometrically consistent details. We evaluate our method for synthetic and real-world datasets, and perform complex spatial properties and its effectiveness in capturing visual dynamics. Experimental results indicate that NeRF-Cap outperforms existing captioning models in terms of spatial awareness, contextual accuracy, and natural language fluency, as measured by standard benchmarks such as Bilingual Evaluation Understudy (BLEU), Metric for Evaluation of Translation with Explicit Ordering (METEOR), Consensus-based Image Description Evaluation (CIDEr) and a novel Depth-Awareness Score. Our work highlights the potential of 3D-aware multimodal captioning, paving the way for more advanced applications in robotic perception, augmented reality, and assistive vision systems. 
653 |a Language 
653 |a Accuracy 
653 |a Deep learning 
653 |a Optimization techniques 
653 |a Neural networks 
700 1 |a Parimala, Rajeshwari 
700 1 |a Motukuri, Sridevi 
700 1 |a Raja Sekhar Reddy Pocha 
700 1 |a Pulipati, Srilatha 
773 0 |t Journal Europeen des Systemes Automatises  |g vol. 58, no. 5 (May 2025), p. 1059-1065 
786 0 |d ProQuest  |t ABI/INFORM Global 
856 4 1 |3 Citation/Abstract  |u https://www.proquest.com/docview/3231508339/abstract/embedded/7BTGNMKEMPT1V9Z2?source=fedsrch 
856 4 0 |3 Full Text - PDF  |u https://www.proquest.com/docview/3231508339/fulltextPDF/embedded/7BTGNMKEMPT1V9Z2?source=fedsrch