Post-Processing: NMS Variants and Soft-NMS
What You'll Learn
You'll understand how to improve detection results beyond basic NMS by using soft-scoring methods that preserve more valid detections while still removing duplicates.
The Problem with Standard NMS
You've already learned about Non-Maximum Suppression (NMS) in lesson 954. Standard NMS uses a binary decision: if two bounding boxes overlap beyond a threshold (typically IoU > 0.5), it completely eliminates the lower-confidence box. This works well, but it's aggressive—sometimes it discards valid detections, especially when objects are truly close together or partially occluding each other.
Soft-NMS: A Gentler Approach
Soft-NMS doesn't completely eliminate overlapping boxes. Instead, it reduces their confidence scores based on how much they overlap with higher-scoring boxes. Think of it like this: standard NMS is a light switch (on/off), while Soft-NMS is a dimmer (gradual adjustment).
The key insight: if a box overlaps significantly with a better detection, it's probably a duplicate—but maybe not entirely. Rather than delete it, reduce its confidence proportionally. If it's still confident enough after penalty, keep it around.
How Soft-NMS Works
- Start with the highest-confidence detection (just like NMS)
- For all remaining boxes that overlap with it:
- Instead of deleting them, decay their scores based on IoU
- Higher overlap → larger penalty
- Move to the next highest-confidence box and repeat
- Finally, apply a confidence threshold to filter low-scoring boxes
Common decay functions:
- Linear:
score = score × (1 - IoU) - Gaussian:
score = score × exp(-(IoU² / σ))
When to Use Soft-NMS
Soft-NMS particularly helps when detecting crowded scenes or overlapping objects where standard NMS might be too harsh. It typically improves mAP by 1-2% with minimal computational overhead.
Key Takeaway: Soft-NMS refines post-processing by penalizing rather than eliminating overlapping detections, preserving more valid boxes while still suppressing duplicates—especially valuable in crowded or occluded scenarios.