Few-Shot Object Detection
Few-Shot Object Detection with Contrastive Learning · Also known as: FSOD, Few-shot detection
Few-Shot Object Detection (FSOD) is a meta-learning approach that enables detecting novel object classes from only a few annotated examples. Unlike standard object detection requiring hundreds of labeled instances per class, FSOD learns to quickly adapt detection models to new object categories by leveraging knowledge from base categories.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Few-shot object detection is essential in domains where annotating new object classes is expensive or time-consuming, such as automated inspection of new product variants or detecting rare medical conditions. Use standard object detection when abundant labeled examples are available for all target classes. FSOD is particularly valuable for scenarios where detection models must quickly adapt to emerging categories.
Strengths & limitations
- Enables detection of novel object classes with only 1-10 annotated examples per class
- Leverages pre-training on base classes to learn transferable representations
- Attention mechanisms enable the model to focus on relevant object regions without exhaustive search
- Relation networks naturally handle variable support set sizes
- Performance on novel classes lags significantly behind detectors trained on hundreds of examples
- Requires careful selection of support sets; poor examples lead to degraded performance
- Pre-training data distribution significantly impacts few-shot adaptation; domain mismatch hurts transfer
Frequently asked
What is the difference between few-shot and zero-shot object detection?
Few-shot detection provides a small set of labeled examples (support set) to adapt the model. Zero-shot detection must generalize to novel classes using only class descriptions or semantic embeddings without visual examples. Few-shot detection typically achieves higher accuracy because visual examples contain richer information than text descriptions.
How are support sets selected?
Support sets should contain diverse, representative examples that cover variations in appearance, pose, and scale. Random selection often underperforms; active selection strategies that choose maximally informative examples usually work better. The quality of support examples directly impacts detection performance.
How does pre-training on base classes help?
Pre-training learns a feature space where objects of the same class have similar representations while different classes are well-separated. This learned feature space generalizes to novel classes, enabling the model to recognize new objects despite never seeing them during training. The quality of base class data directly determines transfer effectiveness.
Sources
- Wang, X., Huang, T. E., Darrell, T., Gonzalez, J. E., & Yu, F. (2020). Few-shot object detection with attention-RPN and multi-relation detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 9050-9059). link ↗
How to cite this page
ScholarGate. (2026, June 3). Few-Shot Object Detection with Contrastive Learning. ScholarGate. https://scholargate.app/en/deep-learning/few-shot-object-detection
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- DETR (Detection Transformer)Deep learning↔ compare
- SimCLRDeep learning↔ compare
- Swin TransformerDeep learning↔ compare