Abstract
We present FOCAL, an interactive lifelog retrieval system for the 9th Lifelog Search Challenge (LSC’26), designed to improve the retrieval of fine-grained visual details. While vision-language models such as CLIP provide strong global representations, they often overlook small or partially visible objects that are critical for resolving lifelog queries. To address this, FOCAL introduces an object-centric attention mechanism that integrates YOLO-World open-vocabulary detection with FALIP-guided attention, enabling region-aware embedding refinement while preserving global context. The system further incorporates a temporal alignment module based on the K-pointer algorithm to support efficient multi-event retrieval. A unified Elasticsearch index enables hybrid search by combining vector similarity with structured metadata and OCR-based text filtering. Additionally, an interactive interface supports flexible query construction with configurable temporal constraints. Preliminary results show that FOCAL improves retrieval accuracy for object-focused queries while maintaining low latency for interactive use, demonstrating the effectiveness of attention-guided retrieval in lifelog scenarios.
Authors
Duc Anh Khoa Dinh, Duc-Tai Dinh, Minh-Quan Ho-Le, Van-Tu Ninh, Minh-Triet Tran
Highlights
- Object-centric retrieval: Combines the YOLO-World v2 object detection model and the FALIP attention-orientation mechanism to highlight image regions containing important objects while preserving the global context.
- Hybrid search: Combines vector similarity (CLIP and FALIP embeddings) with structural metadata filters (OCR, time, location, activity) on the Elasticsearch platform.
- Temporal alignment: Applying the improved K-pointer algorithm to effectively search for event sequences in chronological order.
- Optimal interactive interface: The interface supports flexible query structures, allows adjusting time resolution and displays results visually in groups.
- Top 6 LSC’26. KIS: 712.2; AVS: 849.9; QA: 661.9; Total: 2224.0
References
[ACM] FOCAL: Focused Object-Centric Attention for Lifelog Retrieval
[LSC’26] FOCAL: Focused Object-Centric Attention for Lifelog Retrieval. Presenter: DucAnh Khoa Dinh
Lifelog Search Challenge - LSC’26. LSC’26 took place on June 16, 2026 at ICMR’26.
BibTeX
@inproceedings{dinh2026focal, author = {Dinh, Duc Anh Khoa and Dinh, Duc-Tai and Ho-Le, Minh-Quan and Ninh, Van-Tu and Tran, Minh-Triet}, title = {FOCAL: Focused Object-Centric Attention for Lifelog Retrieval}, booktitle = {Proceedings of the 9th Annual Lifelog Search Challenge (LSC '26)}, year = {2026}, pages = {to be published}, address = {Amsterdam, The Netherlands}, publisher = {ACM}}
