Weakly Supervised Object Detection
Weakly Supervised Object Detection (WSOD) · Also known as: WSOD, weakly-supervised detection, image-level supervised detection, multiple instance detection
Weakly Supervised Object Detection (WSOD) trains object detectors using only image-level labels — indicating which object classes appear in an image — without requiring costly bounding-box annotations. Multiple Instance Learning (MIL) formulations allow the model to discover the likely location of each object class from classification signals alone, dramatically reducing annotation cost.
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
Use WSOD when bounding-box annotations are prohibitively expensive or unavailable and only image-level labels can be collected at scale — for example, in medical imaging, satellite imagery, or niche industrial domains. It is appropriate when approximate localization is acceptable and a moderate drop in detection accuracy relative to fully supervised methods is tolerable. Do not use WSOD when precise bounding-box localization is required, when the target domain has small, densely packed, or heavily occluded objects (where MIL assumptions break down), or when a modest fully annotated dataset can be acquired — fully supervised detection will consistently outperform WSOD in such cases.
Strengths & limitations
- Eliminates the need for expensive bounding-box annotation, lowering data collection cost by an order of magnitude.
- Enables object detection in domains where box-level labels are impractical to obtain at scale.
- Compatible with standard CNN and Transformer backbones pre-trained on large image corpora.
- Progressive refinement strategies (e.g., OICR, PCL) substantially close the gap to fully supervised baselines.
- Can be combined with semi-supervised techniques when a small number of box annotations is available.
- Consistently underperforms fully supervised detectors, often by a large margin on standard benchmarks.
- MIL assumptions fail for small objects, occluded objects, and scenes with many co-occurring instances of the same class.
- Training is sensitive to the quality and diversity of region proposals; poor proposals lead to systematic localization failures.
- Pseudo-label noise accumulates across refinement iterations, potentially reinforcing localization errors.
- Results are hard to reproduce across implementations due to sensitivity to proposal parameters and training schedules.
Frequently asked
How does WSOD differ from semi-supervised object detection?
WSOD uses no bounding-box annotations at all — only image-level labels. Semi-supervised object detection uses a small set of fully annotated (box-level) images together with a larger pool of unannotated or weakly annotated images.
Why is WSOD so much harder than weakly supervised classification?
Classification only needs to identify that a class is present; detection must also localize it accurately. The MIL bag assumption becomes fragile when multiple object instances appear in the same image or objects are small relative to the full image.
Can I use pre-trained models to bootstrap WSOD?
Yes. Starting from an ImageNet-pre-trained backbone substantially improves region feature quality. Vision-language models such as CLIP also provide strong image-level supervision signals that can guide weakly supervised localization.
What benchmark datasets are standard for WSOD?
Pascal VOC 2007 and VOC 2012 are the most commonly reported benchmarks; COCO is used for harder evaluations. mAP at IoU 0.5 is the primary metric, and CorLoc (correct localization rate) is used to evaluate localization quality separately.
When should I collect a small set of bounding boxes instead of using WSOD?
If you can annotate even a few hundred boxes per class, a semi-supervised or few-shot detection approach will typically outperform a purely weakly supervised pipeline. Reserve WSOD for situations where box annotation is structurally impossible, not merely inconvenient.
Sources
- Bilen, H., & Vedaldi, A. (2016). Weakly supervised deep detection networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2846–2854. DOI: 10.1109/CVPR.2016.311 ↗
- Tang, P., Wang, X., Bai, X., & Liu, W. (2017). Multiple instance detection network with online instance classifier refinement. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2843–2851. DOI: 10.1109/cvpr.2017.326 ↗
How to cite this page
ScholarGate. (2026, June 3). Weakly Supervised Object Detection (WSOD). ScholarGate. https://scholargate.app/en/deep-learning/weakly-supervised-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.
- Image ClassificationDeep learning↔ compare
- Instance SegmentationDeep learning↔ compare
- Object DetectionDeep learning↔ compare
- Semi-supervised Object DetectionDeep learning↔ compare
- Vision TransformerDeep learning↔ compare