OBJECT DETECTION WITH GROUNDING DINO: SWIN-T AND INDOBERT FOR INDONESIAN TEXT-IMAGE-MATCHING - Dalam bentuk buku karya ilmiah
DIVA ANINDITHA
Informasi Umum
Kode
25.05.675
Klasifikasi
000 - General Works
Jenis
Karya Ilmiah - Thesis (S2) - Reference
Subjek
Artificial Intelegence
Dilihat
61 kali
Informasi Lainnya
Abstraksi
The development of open-vocabulary object detection (OVOD) has enabled object recognition from free-form textual prompts. However, most existing OVOD systems, including Grounding DINO, remain constrained to English-language datasets and encoders, limiting their applicability in non-English contexts. This research explores the adaptation of Grounding DINO by replacing its default BERT text encoder with IndoBERT, a monolingual language model trained on Indonesian corpora. Using manually translated subset of the COCO val2017 dataset, this study evaluates the performance of three model configurations: (1) English captions with BERT, (2) Indonesian captions with BERT, and (3) Indonesian captions with IndoBERT. The IndoBERT-enhanced model achieved a precision of 0.758, F1-score of 0.21, and [email protected]:0.9 of 0.132, outperforming the baselines in aligning Indonesian prompts with visual objects. These findings support the feasibility of developing vision-language models tailored to low-resource languages, emphasizing the role of monolingual encoders in cross-modal alignment.
- ABK7ZAA5 - Tesis 2
Koleksi & Sirkulasi
Tersedia 1 dari total 1 Koleksi
Anda harus log in untuk mengakses flippingbook
Pengarang
| Nama | DIVA ANINDITHA |
| Jenis | Perorangan |
| Penyunting | Suryo Adhi Wibowo, Koredianto Usman |
| Penerjemah |
Penerbit
| Nama | Universitas Telkom, S2 Teknik Elektro |
| Kota | Bandung |
| Tahun | 2025 |
Sirkulasi
| Harga sewa | IDR 0,00 |
| Denda harian | IDR 0,00 |
| Jenis | Non-Sirkulasi |
