Post-Processing

uq_detr.select(detections, method, param)

Select a subset of detections via common post-processing strategies. This is the bridge between raw model output (hundreds of queries) and the filtered set used for evaluation.

Methods

Method	param meaning	Description
"threshold"	Confidence threshold	Keep detections with max confidence > param
"topk"	Number of detections	Keep the top-k by max confidence
"nms"	IoU threshold	Non-maximum suppression with the given IoU threshold

Usage

from uq_detr import select

# Confidence thresholding
filtered = select(all_queries, method="threshold", param=0.3)

# Top-k selection (e.g., DINO uses top-300 out of 900)
filtered = select(all_queries, method="topk", param=300)

# Non-maximum suppression
filtered = select(all_queries, method="nms", param=0.5)

Comparing Post-Processing Strategies

A key use case: sweep configurations and use OCE to find the best one.

import uq_detr
from uq_detr import select
import numpy as np

# Threshold sweep
for thr in np.arange(0.1, 0.9, 0.1):
    filtered = [select(q, method="threshold", param=thr) for q in all_queries]
    print(f"thr={thr:.1f}  OCE={uq_detr.oce(filtered, gts).score:.4f}")

# Top-k sweep
for k in [10, 50, 100, 300]:
    filtered = [select(q, method="topk", param=k) for q in all_queries]
    print(f"top-{k}  OCE={uq_detr.oce(filtered, gts).score:.4f}")