Implementation:Obss Sahi Legacy NMS Postprocess
| Knowledge Sources | |
|---|---|
| Domains | Object_Detection, Postprocessing |
| Last Updated | 2026-02-08 12:00 GMT |
Overview
⚠️ DEPRECATED: Legacy pure-Python implementations of Non-Maximum Suppression (NMS) and Union-Merge postprocessing for object detection predictions. Located in sahi/postprocess/legacy/ directory. Superseded by sahi.postprocess.combine.
Description
⚠️ This module is deprecated. It resides in the legacy/ directory and has been superseded by the optimized implementations in sahi.postprocess.combine. See Heuristic:Obss_Sahi_Warning_Deprecated_Legacy_NMS for migration guidance.
The sahi.postprocess.legacy.combine module provides the original Python-based postprocessing classes that preceded the current optimized implementations in sahi.postprocess.combine. It contains three classes:
- PostprocessPredictions — Base class providing IOU/IOS match calculation, configurable match threshold, and class-agnostic/per-class toggle.
- NMSPostprocess — Standard greedy NMS: iteratively selects the highest-confidence prediction and removes all overlapping candidates above the match threshold.
- UnionMergePostprocess — Instead of discarding overlapping predictions, merges their bounding boxes (union), scores (max), and masks (logical OR).
These classes operate on ObjectPrediction instances and use numpy for geometry calculations.
Usage
Use these legacy classes only when backward compatibility with older SAHI versions is required. For new code, prefer the optimized implementations in sahi.postprocess.combine.
Code Reference
Source Location
- Repository: Obss_Sahi
- File: sahi/postprocess/legacy/combine.py
- Lines: 1-179
Signature
class PostprocessPredictions:
def __init__(
self,
match_threshold: float = 0.5,
match_metric: str = "IOU",
class_agnostic: bool = True,
): ...
class NMSPostprocess(PostprocessPredictions):
def __call__(
self,
object_predictions: list[ObjectPrediction],
) -> list[ObjectPrediction]: ...
class UnionMergePostprocess(PostprocessPredictions):
def __call__(
self,
object_predictions: list[ObjectPrediction],
) -> list[ObjectPrediction]: ...
Import
from sahi.postprocess.legacy.combine import NMSPostprocess, UnionMergePostprocess
I/O Contract
Inputs
| Name | Type | Required | Description |
|---|---|---|---|
| match_threshold | float | No (default 0.5) | IOU/IOS threshold for considering two predictions as overlapping |
| match_metric | str | No (default "IOU") | Overlap metric: "IOU" (intersection over union) or "IOS" (intersection over smaller) |
| class_agnostic | bool | No (default True) | If True, match across different category IDs |
| object_predictions | list[ObjectPrediction] | Yes | Detection predictions to postprocess |
Outputs
| Name | Type | Description |
|---|---|---|
| selected_object_predictions | list[ObjectPrediction] | Filtered/merged predictions after NMS or union-merge |
Usage Examples
Legacy NMS
from sahi.postprocess.legacy.combine import NMSPostprocess
# Create NMS postprocessor with IOU threshold 0.5
nms = NMSPostprocess(match_threshold=0.5, match_metric="IOU", class_agnostic=True)
# Apply to list of ObjectPrediction instances
filtered_predictions = nms(object_predictions)
Legacy Union Merge
from sahi.postprocess.legacy.combine import UnionMergePostprocess
# Create union-merge postprocessor using IOS metric
merge = UnionMergePostprocess(match_threshold=0.6, match_metric="IOS", class_agnostic=False)
# Apply to list of ObjectPrediction instances
merged_predictions = merge(object_predictions)