EW-DETR: Evolving World Object Detection via Incremental Low-Rank DEtection TRansformer

Munish Monga; Vishal Chudasama; Pankaj Wasnik; C. V. Jawahar

cs.CVFeb 24, 2026

EW-DETR: Evolving World Object Detection via Incremental Low-Rank DEtection TRansformer

Authors: Munish Monga, Vishal Chudasama, Pankaj Wasnik, C. V. Jawahar

Organizations: Sony Research India · IIIT Hyderabad

Abstract

Real-world object detection must operate in evolving environments where new classes emerge, domains shift, and unseen objects must be identified as "unknown": all without accessing prior data. We introduce Evolving World Object Detection (EWOD), a paradigm coupling incremental learning, domain adaptation, and unknown detection under exemplar-free constraints. To tackle EWOD, we propose EW-DETR framework that augments DETR-based detectors with three synergistic modules: Incremental LoRA Adapters for exemplar-free incremental learning under evolving domains; a Query-Norm Objectness Adapter that decouples objectness-aware features from DETR decoder queries; and Entropy-Aware Unknown Mixing for calibrated unknown detection. This framework generalises across DETR-based detectors, enabling state-of-the-art RF-DETR to operate effectively in evolving-world settings. We also introduce FOGS (Forgetting, Openness, Generalisation Score) to holistically evaluate performance across these dimensions. Extensive experiments on Pascal Series and Diverse Weather benchmarks show EW-DETR outperforms other methods, improving FOGS by 57.24%.

Figures & tables

Figure 2: Top Row: gives an overview of the EW-DETR framework, which augments the standard DETR pipeline with three key modules: Incremental LoRA Adapters for exemplar-free incremental learning under evolving domains, Query-Norm Objectness Adapter for decoupled objectness-aware features, and Entropy-Aware Unknown Mixing for calibrated unknown detection. Bottom Row: illustrates a quick qualitative comparison of EWOD with other OD approaches. Best viewed in colour with zoom 1 1 1 Images are generated using Gemini 2.5 Pro for illustration purposes. .

Figure 3: EW-DETR framework for Evolving-World Object Detection. An input image is processed by a frozen backbone and transformer encoder–decoder equipped with aggregate $\big(\Delta\mathbf{W}^{t-1}_{\text{agg}}\big)$ and task-specific $\big(\Delta\mathbf{W}^{t}_{\text{task}}\big)$ LoRA adapters, yielding class-agnostic query features. These queries are reparameterised by the Query-Norm Objectness Adapter and passed through a classification head and an objectness head, whose outputs are combined by Entropy-aware Unknown Mixing module to produce calibrated class scores, while a localisation head predicts bounding boxes to form final detections across tasks 1 1 footnotemark: 1 .

Method	Trainable Params (M)	$\mathcal{T}_{1}$ : VOC [1:10]		$\mathcal{T}_{2}$ : Clipart [11:18]				Metrics
			mAP ( $\uparrow$ )		mAP ( $\uparrow$ )
		U-Recall ( $\uparrow$ )	Curr. Known	U-Recall ( $\uparrow$ )	Prev. Known	Curr. Known	Both	FSS ( $\uparrow$ )	OSS ( $\uparrow$ )	GSS ( $\uparrow$ )	FOGS ( $\uparrow$ )
ORE-EBUI ${}_{{\color[rgb]{0,0.44,0.57}\text{ CVPR'}21}}$ [ 17 ]	32.96	9.84	23.7	6.97	0	11.37	5.05	0	55.48	11.37	22.28
OW-DETR ${}_{{\color[rgb]{0,0.44,0.57}\text{ CVPR'}22}}$ [ 11 ]	24.22	16.25	31.51	8.07	3.6	7.96	5.54	11.42	40.47	7.96	19.95
PROB ${}_{{\color[rgb]{0,0.44,0.57}\text{ CVPR'}23}}$ [ 48 ]	23.99	52.73	12.68	46.27	0	0.27	0.12	0	67.58	0.27	22.62
CAT ${}_{{\color[rgb]{0,0.44,0.57}\text{ CVPR'}23}}$ [ 23 ]	24.25	20.23	30.18	8.1	8.91	8.05	8.53	29.52	41.29	8.05	26.29

Table 1: EWOD on Pascal Series: VOC [1:10]

\to

Clipart [11:18]. Along with proposed metrics, per-task OWOD metrics: U-Recall and [email protected] are also reported for completeness. Best results per column in bold , second-best underlined .

Method	VOC [1:10] $\to$ Watercolor [11:14]				VOC [1:10] $\to$ Comic [11:14]				Daytime Sunny [1:2] $\to$ Night Sunny [3:4] $\to$ Night Rainy [5:6]				Daytime Sunny [1:2] $\to$ Daytime Foggy [3:4] $\to$ Dusk Rainy [5:6]
Method	FSS	OSS	GSS	FOGS	FSS	OSS	GSS	FOGS	FSS	OSS	GSS	FOGS	FSS	OSS	GSS	FOGS
ORE-EBUI [ 17 ]	0	50.7	21.46	24.05	0	52.34	18.25	23.53	0	61.51	18.96	26.82	0	61.14	24.73	28.62
OW-DETR [ 11 ]	11.9	41.57	10.16	21.21	1.43	41.83	8.75	17.34	0.76	61.95	21	27.9	2.31	63.02	22.23	29.19
PROB [ 48 ]	0	63.98	3.45	22.48	0	63.75	4.06	22.6	1.24	66.25	14.8	27.43	0	66.9	12.35	26.42
CAT [ 23 ]	1.51	44.03	12.57	23.7	7.82	43	10.31	20.38	4.01	62.92	23.66	30.2	4.85	63.87	20.5	29.74
ORTH [ 36 ]	12.32	46.89	45.01	34.74	5.28	51.56	38.49	31.78	0.08	62.68	40.38	34.38	0.31	61.81	48.56	36.89

Table 2: EWOD results on Pascal Series (left: two-phase) and Diverse Weather (right: multi-phase). Best in bold , second-best underlined .

Method	Daytime Sunny [1:3] $\to$ Night Sunny [4:6]				Daytime Sunny [1:3] $\to$ Night Rainy [4:6]				Daytime Sunny [1:3] $\to$ Daytime Foggy[4:6]				Daytime Sunny [1:3] $\to$ Dusk Rainy [4:6]
Method	FSS	OSS	GSS	FOGS	FSS	OSS	GSS	FOGS	FSS	OSS	GSS	FOGS	FSS	OSS	GSS	FOGS
ORE-EBUI [ 17 ]	0	61.06	20.41	27.16	0	61.44	9.18	23.54	0	60.35	24.36	28.24	0	61.51	16.45	25.99
OW-DETR [ 11 ]	0	61.06	23	28.02	6.54	61.79	6.84	25.06	0	61.49	23.39	28.29	0.71	60.39	12.47	24.52
PROB [ 48 ]	0	69.01	8.25	25.75	0.96	69.35	3.98	24.76	0	69.49	10.57	26.69	0	69.15	11.34	26.83
CAT [ 23 ]	0	60.77	24.43	28.4	7.71	61.47	7.9	25.69	0	60.72	29.02	29.91	1.1	60.85	12.41	24.56
ORTH [ 36 ]	0.31	59.33	40.56	33.4	2.61	64.21	22.81	29.88	0.54	60.94	52.89	38.12	0.49	67.88	36.31	34.89

Table 3: EWOD on Diverse Weather Series: Daytime Sunny

\to

Night/Foggy/Rainy transitions. Best in bold , second-best underlined .

Figure 4: t-SNE visualisation of decoder query features comparing EW-DETR with recent methods on VOC [1:10] $\to$ Clipart [11:18] task. The top row shows features obtained during inference on VOC [1:10] classes, while the bottom row shows features evaluated on the combined test set comprising VOC [1:18] and Clipart [1:18] classes. Notably, EW-DETR is only able to preserve distinct class clusters even under severe domain shift (VOC $\to$ Clipart) while also generalising to unseen class-domain pairs: VOC [11:18] and Clipart [1:10], whereas other recent methods exhibit severe feature collapse in $\mathcal{T}_{2}$ . Best viewed in colour with zoom.

Configuration	$\mathcal{T}_{1}$ : VOC [1:10]		$\mathcal{T}_{2}$ : Clipart [11:18]				Metrics				Trainable Params (M)
		mAP ( $\uparrow$ )		mAP ( $\uparrow$ )
	U-Recall ( $\uparrow$ )	Curr. Known	U-Recall ( $\uparrow$ )	Prev. Known	Curr. Known	Both	FSS ( $\uparrow$ )	OSS ( $\uparrow$ )	GSS ( $\uparrow$ )	FOGS ( $\uparrow$ )
Baseline	1.46	77.21	0.56	5.81	51.49	26.11	7.52	33.78	51.49	30.93	30.8672
+ Train $\Delta\mathbf{W}^{t}_{\text{task}}$ only (simple PEFT, no agg.)	1.20	77.23	0.84	20.48	52.33	34.63	26.52	33.34	52.33	37.4	1.7997
+ Train $\Delta\mathbf{W}^{t-1}_{\text{agg}}~\&~\Delta\mathbf{W}^{t}_{\text{task}}$ (dual LoRA; no merging/SVD)	1.77	77.07	0.11	20.59	50.49	33.87	26.72	34.26	50.49	37.16	1.7997
+ Train Incre. LoRA	1.08	76.29	0.29	74.85	0.07	41.61	98.11	33.53	0.07	43.9	1.7997

Table 4: Ablations of proposed modules of the EW-DETR framework on RF-DETR [ 33 ] for: VOC [1:10]

\to

Clipart [11:18] benchmark.

Panel of Figure 8 — Figure S1 : Effect of 1(a) data-aware vs. 1(b) fixed merging.

Figure S2 : Histogram plots of last-layer decoder query norms for VOC [1:10] $\to$ Clipart [11:18].

Daytime Sunny [1:3] $\rightarrow$ Night Sunny [4:6]
Task	Train Classes	Train Domain	Test Classes	Test Domains
$\mathcal{T}_{1}$	Daytime Sunny [1:3] bike, bus, car	Daytime Sunny	Daytime Sunny [1:3] bike, bus, car + unknown	Daytime Sunny
$\mathcal{T}_{2}$	Night Sunny [4:6] motor, person, rider , truck	Night Sunny	Daytime Sunny [1:6] + Night Sunny [1:6] bike, bus, car + motor, person, rider + unknown	Daytime Sunny + Night Sunny
Daytime Sunny [1:2] $\rightarrow$ Night Sunny [3:4] $\rightarrow$ Night Rainy [5:6]
Task	Train Classes	Train Domain	Test Classes	Test Domains
$\mathcal{T}_{1}$	Daytime Sunny [1:2] bike, bus	Daytime Sunny	Daytime Sunny [1:2] bike, bus + unknown	Daytime Sunny

Table S1 : EWOD Training and Evaluation Protocol in Diverse Weather benchmark. Blue , orange , and olive denote classes introduced in

\mathcal{T}_{1}

\mathcal{T}_{2}

, and

\mathcal{T}_{3}

, respectively; red indicates classes excluded during training in all tasks but treated as unknown during testing.

Class ID	Class Name	Training			Testing
Class ID	Class Name	$\mathcal{T}_{1}:$ Daytime Sunny	$\mathcal{T}_{2}:$ Night Sunny	$\mathcal{T}_{3}:$ Night Rainy	$\mathcal{T}_{1}:$ Daytime Sunny	$\mathcal{T}_{2}:$ Daytime Sunny + Night Sunny	$\mathcal{T}_{3}:$ Daytime Sunny + Night Sunny + Night Rainy
1	bike	✓			✓	✓ ✓	✓ ✓ ✓
2	bus	✓			✓	✓ ✓	✓ ✓ ✓
3	car		✓			✓ ✓	✓ ✓ ✓
4	motor		✓			✓ ✓	✓ ✓ ✓
5	person			✓			✓ ✓ ✓

Table S2 : Class-wise distribution on Diverse Weather benchmark: Daytime Sunny [1:2]

\to

Night Sunny [3:4]

\to

Night Rainy [5:6]; red indicates classes (i.e., truck ) excluded during training in all tasks but treated as unknown during testing. Best viewed in colour.

Class ID	Class Name	Training				Testing
Class ID	Class Name	$\mathcal{T}_{1}:$ VOC	$\mathcal{T}_{2}:$ Clipart	$\mathcal{T}_{2}:$ Watercolor	$\mathcal{T}_{2}:$ Comic	$\mathcal{T}_{1}:$ VOC	$\mathcal{T}_{2}:$ VOC + Clipart	$\mathcal{T}_{2}:$ VOC + Watercolor	$\mathcal{T}_{2}:$ VOC + Comic
1	aeroplane	✓				✓	✓ ✓	✓	✓
2	boat	✓				✓	✓ ✓	✓	✓
3	bottle	✓				✓	✓ ✓	✓	✓
4	bus	✓				✓	✓ ✓	✓	✓
5	chair	✓				✓	✓ ✓	✓	✓

Table S3 : Class-wise distribution on three Pascal Series benchmarks:

\mathcal{T}_{1}:

VOC [1:10]

\to

\mathcal{T}_{2}:

{Clipart [11:18] / Watercolor [11:14] / Comic [11:14]}. Note that while Watercolor and Comic classes are listed here as IDs 15–18 for visual continuity, they are strict subsets of the whole label space and are mapped to IDs 11–14 during experiments; red indicates classes (i.e., dog , person ) excluded during training in all tasks but treated as unknown during testing. Best viewed in colour.

Method	Underlying Detector	Trainable Params (M)	GFLOPs	Avg. Inference Speed (ms)	Avg. Memory Footprint (GB)
ORE [ 17 ]	Faster RCNN [ 32 ]	32.96	1665.03	60.41	3.75
OW-DETR [ 11 ]	D-DETR [ 47 ]	24.22	157.71	59.82	1.33
PROB [ 48 ]	D-DETR [ 47 ]	23.99	135.31	97.71	1.32
CAT [ 23 ]	D-DETR [ 47 ]	24.25	135.34	112.54	1.35
ORTH [ 36 ]	RandBox [ 40 ]	105.9	2073.57	1151.19	1.26
DuET [ 26 ]	D-DETR [ 47 ]	24.22	135.3	150.61	1.41

Table S4 : Computational complexity analysis of various methods evaluated on Pascal Series (VOC [1:10]

\to

Clipart [11:18]) benchmark, all evaluated on a single NVIDIA H100 80GB card.

$\mathcal{T}_{1}$	$\mathcal{T}_{2}$	$\mathcal{T}_{3}$	FSS	OSS	GSS	FOGS
Daytime Sunny [5:6]	Night Sunny [1:2]	Night Rainy [3:4]	73.30	73.45	24.62	57.12
Daytime Sunny [3:4]	Night Sunny [5:6]	Night Rainy [1:2]	75.03	71.31	17.94	54.76
Daytime Sunny [1:2]	Night Sunny [3:4]	Night Rainy [5:6]	73.63	73.43	18.68	55.25
Night Rainy [1:2]	Daytime Sunny [3:4]	Night Sunny [5:6]	71.79	72.50	28.67	57.65
Night Sunny [1:2]	Night Rainy [3:4]	Daytime Sunny [5:6]	71.93	70.69	21.25	54.62
Daytime Sunny [7,2]	Night Sunny [3,4]	Night Rainy [5,6]	73.65	71.71	14.96	53.44

Table S5 : Sensitivity analysis for random task permutations on Diverse Weather benchmark.

Method	$\mathcal{T}_{1}$ : VOC [1:10]			$\mathcal{T}_{2}$ : Clipart [11:18]			Metrics
Method	U-Recall ( $\uparrow$ )	WI ( $\downarrow$ )	A-OSE ( $\downarrow$ )	U-Recall ( $\uparrow$ )	WI ( $\downarrow$ )	A-OSE ( $\downarrow$ )	FSS ( $\uparrow$ )	OSS ( $\uparrow$ )	GSS ( $\uparrow$ )	FOGS ( $\uparrow$ )
ORE-EBUI [ 17 ]	9.84	0.0985	12916	6.97	0.0592	852	0	55.48	11.37	22.28
OW-DETR [ 11 ]	16.25	0.0851	55228	8.07	0.0644	28689	11.42	40.47	7.96	19.95
PROB [ 48 ]	52.73	0.1603	12700	46.27	0.0216	1529	0	67.58	0.27	22.62
CAT [ 23 ]	20.23	0.083	57094	8.1	0.0566	28045	29.52	41.29	8.05	26.29
ORTH [ 36 ]	63.92	0.757	32519	41.61	0.0213	5160	5.83	51.06	32.44	29.78

Table S6 : Analysis of unknown object confusion on Pascal Series benchmark: VOC [1:10]

\to

Clipart [11:18]. The table compares all methods using unknown class confusion metrics, including U-Recall, WI and A-OSE. Best results per column in bold , second-best underlined .

Figure S3 : Qualitative comparison of EW-DETR with other methods on: VOC [1:10] $\to$ Clipart [11:18]. Best viewed in colour with zoom.

Figure S4 : Comprehensive comparison of various methods across all three dimensions of EWOD: Forgetting, Openness, and Generalisation, quantified by FSS, OSS, and GSS, respectively, averaged across all EWOD experiments.

Panel of Figure 18 — Figure S5 : Ablation analysis for key hyperparameters in the EW-DETR framework. 5(a) Sensitivity to LoRA rank $r$ . 5(b) Sensitivity to data-aware merging bounds $(\beta_{\min},\beta_{\max})$ .

Object Detection

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

See all 605 papers

REAL-OW: Rehearsal-free Open World Object Detection with Low-Rank Adaptation and Dual-Stage Objectness Modeling
Jul 3, 2026Huazhong Zhang, Xiaowen Fu, Yang Zhang +2Open-Vocabulary Object Detection Continual Learning
Shenzhen University, Shenzhen, China
Knowing the Unknown: Interpretable Open-World Object Detection via Concept Decomposition Model
Feb 24, 2026Xueqiang Lv, Shizhou Zhang, Yinghui Xing +3Explainable Object Detection Object Detection
School of Computer Science, Northwestern Polytechnical Unviersity, Xi’an, China · Huawei Technologies Ltd.
Detecting Unknown Objects via Energy-based Separation for Open World Object Detection
Mar 31, 2026Jun-Woo Heo, Keonhee Park, Gyeong-Moon ParkContinual Object Detection Object Detection
Korea University, South Korea · Seoul National University, South Korea

Jul 3, 2026J/K move · Enter open · S save

REAL-OW: Rehearsal-free Open World Object Detection with Low-Rank Adaptation and Dual-Stage Objectness Modeling

Huazhong Zhang, Xiaowen Fu, Yang Zhang, Linlin Shen+1

Shenzhen University, Shenzhen, China

Open-Vocabulary Object Detection Continual Learning Object Detection

Open paper AskPDF arXiv

Open-World Object Detection (OWOD) requires detectors to identify previously unseen objects as unknown and incrementally incorporate them into the set of known categories, while preserving previously acquired knowledge. Existing frameworks rely heavily on exemplar replay to mitigate catastrophic forgetting, but in some real applications, storing raw data conflicts with data access restrictions and leads to data exposure risks, while incurring significant memory overhead. In this paper, we propose REAL-OW, a novel rehearsal-free framework that decouples incremental knowledge through a collaborative adapter architecture based on Low-Rank Adaptation (LoRA). Specifically, we deploy General Adapters (GAs) in the backbone to enable the significance-aware refinement of cross-task universal representations, while Specific Adapters (SAs) in the decoder provide orthogonal storage for task-specific expertise. To resolve representation drift in objectness modeling under rehearsal-free constraints, we introduce Dual-Stage Objectness Modeling (DSOM), which alternates between feature aggregation and boundary consolidation to stabilize objectness distributions while maintaining the separation between known and unknown categories. Furthermore, DSOM is supported by a Calibrated Gaussian Negative Log-Likelihood (CG-NLL) distance tailored for the dispersed feature distributions inherent in rehearsal-free settings. Extensive evaluations demonstrate that REAL-OW achieves state-of-the-art performance, surpassing existing exemplar replay methods in both detection precision and unknown discovery. Our approach establishes a new baseline for rehearsal-free OWOD.

Abstract

Object Detection

Explore similar work

REAL-OW: Rehearsal-free Open World Object Detection with Low-Rank Adaptation and Dual-Stage Objectness Modeling

Knowing the Unknown: Interpretable Open-World Object Detection via Concept Decomposition Model

Detecting Unknown Objects via Energy-based Separation for Open World Object Detection