TRIM: Temporal Run-Time Monitoring of 3D Object Detection via Activation Statistics

Hakan Yekta Yatbaz; Konstantinos Koufos; Mehrdad Dianati

cs.CVOct 28, 2026

TRIM: Temporal Run-Time Monitoring of 3D Object Detection via Activation Statistics

Authors: Hakan Yekta Yatbaz, Konstantinos Koufos, Mehrdad Dianati

Organizations: School of Electronics, Electrical Engineering and Computer Science, Queen’s University Belfast, Belfast, UK

Abstract

Reliable perception is essential to intelligent transportation systems, yet deep neural networks can silently miss safety-critical objects while continuing to produce apparently plausible outputs. Existing run-time monitors for 3D object detection generally assess each frame independently and therefore cannot exploit the temporal evolution of occlusion, point sparsity, and scene complexity. We present TRIM, a lightweight temporal monitor that detects missed-object failures without modifying or retraining the underlying detector. TRIM summarises intermediate activations using compact channel-wise statistics and processes short sequences of these summaries with a temporal convolutional network. Experiments on NuScenes show that temporal monitoring improves AUROC from 0.686 to 0.704 and reduces the false alarm rate at 95% recall from 0.818 to 0.759 compared with single-frame monitoring. Early and intermediate backbone features provide the strongest monitoring signal, while the temporal inference head requires only 0.159 ms on a CPU.

Figures & tables

Fig. 1: Overview of TRIM: Intermediate activations from a frozen CenterPoint detector are summarised by TAP-Stats using five per-channel spatial statistics, then L2-normalised and concatenated into per-frame embeddings. A causal TCN processes

T

embeddings to predict the probability of missed detections at the target frame

t+h

, where

h=0

denotes current-frame monitoring.

Method	AUROC $\uparrow$	$R_{+}\uparrow$	$R_{-}\uparrow$	FPR $\downarrow$	FAR ${}_{95}\downarrow$
Random	0.500	0.500	0.500	0.500	—
Confidence	0.436	0.984	0.010	0.990	0.957
Temporal conf.	0.423	0.949	0.042	0.958	0.959
$k$ -NN ( $k{=}50$ )	0.617	0.661	0.495	0.505	0.874
Full TAP, $T{=}1$	0.679 $\pm$ 0.001	0.708 $\pm$ 0.106	0.537 $\pm$ 0.123	0.463 $\pm$ 0.123	0.826 $\pm$ 0.005
b0+b1, $T{=}1$	0.686 $\pm$ 0.001	0.735 $\pm$ 0.083	0.527 $\pm$ 0.080	0.473 $\pm$ 0.080	0.818 $\pm$ 0.009

TABLE I: Test set results (mean

\pm

std over 3 seeds). Learned monitors use a 0.5 threshold; baseline thresholds maximise validation balanced accuracy. Best learned results in bold.

Fig. 2: False-alarm rate (FAR) versus target true-positive rate (TPR) on the test set. Curves show empirical ROC operating points evaluated on a one-percentage-point target TPR grid; the nine TPR targets reported in the tables are a subset of these points. Shading indicates one population standard deviation across three seeds for learned monitors. The dashed diagonal represents chance performance.

Temporal window $T$ (at $h{=}0$ )
Setting	AUROC $\uparrow$	$R_{+}\uparrow$	$R_{-}\uparrow$	FAR ${}_{95}\downarrow$
$T{=}1$ (frame-level)	0.679	0.708	0.537	0.826
$T{=}2$ ( $0.5$ s)	0.694	0.596	0.671	0.798
$T{=}4$ ( $1.5$ s)	0.692	0.567	0.708	0.827
$T{=}6$ ( $2.5$ s)	0.697	0.608	0.672	0.781
$T{=}8$ ( $3.5$ s)	0.697	0.477	0.767	0.774

TABLE II: Temporal ablations on the test set (full TAP,

\delta=3

m, 3 seeds). Upper: window

T

h=0

; lower: horizon

h

T=4

Taps used	Dim	AUROC $\uparrow$	$R_{+}\uparrow$	$R_{-}\uparrow$	FAR ${}_{95}\downarrow$
Block 0	640	0.661	0.597	0.607	0.838
Block 1	1280	0.694	0.667	0.608	0.805
Neck	1280	0.637	0.561	0.648	0.898
b0 + b1	1920	0.704	0.661	0.630	0.759
b0 + neck	1920	0.667	0.604	0.615	0.849
b1 + neck	2560	0.685	0.804	0.438	0.831

TABLE III: TAP layer ablation on the test set (

T=4

h=0

\delta=3

m, 3 seeds). Best result in bold.

Variant	Dim	Params	FLOPs	Latency (ms)
b0+b1, $T{=}1$	1920	418.8K	0.42M	0.107
b0+b1, $T{=}4$	1920	418.8K	1.65M	0.159
3-block, $T{=}4$	3200	582.7K	2.31M	0.174
3-block, $T{=}8$	3200	582.7K	4.61M	0.188

TABLE IV: Standalone complexity of the temporal monitor head. Latency is median CPU inference time, excluding activation extraction and TAP-Stats encoding.

Object Detection

Object detection locates and classifies objects in images or video, balancing localization accuracy, robustness, and inference speed.

See all 629 papers

MAGneT-3D: Monocular and Domain-Generalizable Temporal 3D Detection
Aug 14, 2026Mohamed Kotb, Johannes Meier, Christoph Reich +33D Object Detection Monocular 3D Object Detection
TU Munich · ETH Zurich · MCML +4
Temporal Preservation over Processing: Diagnosing and Designing Spatiotemporal Single-Stage Video Detectors
Jun 30, 2026Karam Tomotaki-Dawoud, Anna Hilsmann, Peter Eisert +1Video Object Detection Real-Time Object Detection
Fraunhofer HHI, Berlin, Germany
LeAD-M3D: Leveraging Asymmetric Distillation for Real-Time Monocular 3D Detection
Dec 5, 2025Johannes Meier, Jonathan Michel, Oussema Dhaouadi +7Efficient Object Detection Real-Time Object Detection
DeepScenario · ETH Zurich · TU Munich +5

Aug 14, 2026J/K move · Enter open · S save

MAGneT-3D: Monocular and Domain-Generalizable Temporal 3D Detection

Mohamed Kotb, Johannes Meier, Christoph Reich, Oussema Dhaouadi+2

TU Munich · ETH Zurich · MCML+4

3D Object Detection Monocular 3D Object Detection Domain Generalization

Open paper AskPDF arXiv

Monocular temporal 3D detection aims to detect objects in 3D, given a monocular video. Query-based 3D detectors unify detection and cross-view association, but their learnable queries fit the spatial distribution of the training data (e.g., field-of-view). We show that this issue is especially severe when these models are applied to monocular video, hindering generalization to unseen datasets and environments. To address this limitation, we introduce MAGneT-3D, the first method for domain-generalized monocular temporal 3D object detection. Instead of relying on static learnable queries, we propose a Domain-Robust Anchor Generator (DRAG) approach that adaptively derives 3D proposals during inference. To further enable domain generalization, we propose a Temporal Refinement and Identity Merging (TRIM) strategy, reducing dependence on specific 3D proposals. To enable comprehensive domain-generalization evaluation, we establish a cross-dataset benchmark spanning nuScenes, Waymo, Lyft, and ONCE. Under zero-shot domain shifts, MAGneT-3D outperforms all baselines, improving NDS from 12.1% to 18.6% while also increasing in-domain accuracy.

Abstract

Object Detection

Explore similar work

MAGneT-3D: Monocular and Domain-Generalizable Temporal 3D Detection

Temporal Preservation over Processing: Diagnosing and Designing Spatiotemporal Single-Stage Video Detectors

LeAD-M3D: Leveraging Asymmetric Distillation for Real-Time Monocular 3D Detection