Brand Logo Detection in Social Media Images
Created a novel dataset and evaluated computer vision models for robust brand logo detection in cluttered social media environments.
TL;DR: Constructed a comprehensive social-media-focused logo detection dataset (VRL-SM-Logo) and demonstrated that models trained on highly diverse social media imagery achieve superior cross-domain generalization compared to standard benchmark datasets.
Technologies & Methodology
- Models & Architecture: DINOv3-ConvNext-Large backbone paired with a DETR (DEtection TRansformer) detection head.
- Algorithms & Statistical Tools: Weighted Box Fusion (WBF) for annotation aggregation, and Ordinary Least Squares (OLS) regression for behavioral analysis.
- Datasets: VRL-SM-Logo (custom dataset: 5,087 images, 9,731 logo instances across 25 brands) and FlickrLogos-27.
- Annotation Tools: CVAT (Computer Vision Annotation Tool).
Key Contributions & Results
- Led the data collection and annotation pipeline, utilizing Weighted Box Fusion (WBF) to generate high-fidelity ground truth bounding boxes from multiple annotations.
- Conducted a systematic quality assurance analysis using OLS regression, revealing that annotator performance degrades significantly during late-night hours or in sessions exceeding 120 minutes.
- Trained a hybrid DINOv3+DETR object detection model to frame logo detection as a robust set-prediction problem.
- Executed a cross-dataset evaluation demonstrating that the model trained on the VRL-SM-Logo dataset suffered only a minimal performance drop (0.07 on AP@0.5:0.95) when tested on out-of-domain data, significantly outperforming the Flickr-trained model’s generalization capabilities.