Evita | Unified RGB-Event Dense Parsing

Concept

Evita starts from a simple observation: RGB frames and event streams are complementary, but they are not naturally aligned in space, time, or representation. The framework therefore addresses dataset alignment, event-format pretraining, and unified RGB-Event feature learning together.

Architecture

The Evita block embeds geometric, spectral, and routing operations inside the encoder itself. This makes RGB-Event interaction an intrinsic part of representation learning rather than a late-stage decoder operation.

Geometric Alignment

Geometric Parallax Rectification predicts bounded offsets and masks to align event features with RGB structure.

Spectral Transfer

Harmonic Spectral Resonance moves event-driven high-frequency cues through the frequency domain while preserving RGB phase information.

Event-Guided Routing

Transient Global Routing uses event features as dynamic queries to gather global photometric context from the RGB stream.

Qualitative and Efficiency Results

Across night, fog, cloudy, rainy, and sunny scenes, Evita produces cleaner dense parsing around thin structures, road boundaries, vehicles, and low-visibility regions. The qualitative comparison highlights that event cues are most useful when RGB alone becomes unreliable.

The accuracy-FLOPs curve further shows why a unified backbone matters: Evita reaches strong mIoU while avoiding the redundant computation of large dual-stream pipelines.

59.57mIoU on DELIVER

80.12mIoU on DDD17

76.80mIoU on DSEC

44.3Mparameters in Evita-L

Qualitative comparison on adverse driving scenes — Figure 3. Qualitative comparison under diverse adverse conditions. Evita preserves semantic structure more consistently than competing RGB-Event fusion methods.

Accuracy versus FLOPs comparison — Figure 4. Accuracy versus computational cost on DELIVER. Evita improves the accuracy-efficiency tradeoff for dense RGB-Event parsing.

Quantitative Comparisons

The tables summarize the reported results while omitting venue labels from the webpage. Evita variants are highlighted in blue, and the best scores are marked in red.

DELIVER dataset
Method	Backbone	Representation	Input size	Params (M)	FLOPs	mIoU (%)
HRFuser	HRFormer-T	EventBinary	1024x1024	49.9	487.5	42.22
TokenFusion	MiT-B2	EventBinary	1024x1024	26.0	188.05	45.63
MultiMAE	ViT-B	EventBinary	1024x1024	100.3	1054.3	57.95
Omnivore	Swin-B	EventBinary	1024x1024	86.7	335.49	56.03
CMX	MiT-B2	EventBinary	1024x1024	66.6	228.8	56.52
CMNeXt	MiT-B2	EventBinary	1024x1024	58.7	233.5	57.48
CMNeXt	MiT-B4	EventBinary	1024x1024	116.5	449.6	58.87
MambaSeg	VMamba-T	EventBinary	1024x1024	25.5	232.45	56.48
OmniSegmentor	DFormer-L	EventBinary	1024x1024	39.0	224.0	58.63
StitchFusion	MiT-B4	EventBinary	1024x1024	65.3	448.5	58.75
Evita-P	Ours-P	EventBinary	1024x1024	1.3	5.2	48.21
Evita-N	Ours-N	EventBinary	1024x1024	4.9	16.2	52.91
Evita-T	Ours-T	EventBinary	1024x1024	6.5	29.8	55.01
Evita-S	Ours-S	EventBinary	1024x1024	21.9	78.2	56.79
Evita-B	Ours-B	EventBinary	1024x1024	34.3	143.7	58.02
Evita-L	Ours-L	EventBinary	1024x1024	44.3	232.1	59.57

DDD17 and DSEC datasets
Method	Backbone	Params	Modality	Rep.	DDD17 FLOPs	DDD17 mIoU	DDD17 Acc	DSEC FLOPs	DSEC mIoU	DSEC Acc
Single Modal Method
SegFormer	MiT-B2	27.5	Image	RGB	16.0	71.05	95.73	67.6	71.99	94.97
SegNeXt	SegNeXt-B	27.6	Image	RGB	10.2	71.46	95.97	40.4	71.55	94.89
EV-SegNet	Xception	34.6	Event	6-Channel	8.0	54.81	89.76	30.6	51.76	88.61
ESS	E2VID	47.2	Event	Voxel Grid	51.8	61.37	91.08	202.6	51.57	89.25
RGB-X Method
TokenFusion	MiT-B2	28.6	Image-Event	Voxel Grid	32.6	67.96	94.77	137.4	70.67	95.20
MultiMAE	ViT-B	89.7	Image-Event	Voxel Grid	56.1	63.31	93.33	291.5	67.11	94.35
HRFuser	HRFormer-T	12.8	Image-Event	Voxel Grid	17.6	73.17	95.41	70.0	59.66	92.76
CMX	MiT-B2	66.6	Image-Event	Voxel Grid	15.7	77.64	96.46	62.1	73.88	95.51
CMX	MiT-B3	106.3	Image-Event	Voxel Grid	23.8	76.83	96.50	94.1	73.24	95.40
CMX	MiT-B4	140.0	Image-Event	Voxel Grid	31.5	77.00	96.36	124.8	74.06	95.54
CMNeXt	MiT-B2	58.7	Image-Event	Voxel Grid	16.0	76.86	96.47	63.4	73.42	95.48
CMNeXt	MiT-B4	116.6	Image-Event	Voxel Grid	31.0	77.70	96.48	122.3	73.84	95.49
CMNeXt	MiT-B5	149.0	Image-Event	Voxel Grid	37.8	77.09	96.47	149.3	74.22	95.59
GeminiFusion	MiT-B2	31.8	Image-Event	Voxel Grid	33.6	71.15	96.15	141.4	71.45	95.31
StitchFusion	MiT-B2	30.0	Image-Event	Voxel Grid	20.8	76.83	96.30	89.7	73.63	95.43
StitchFusion	MiT-B3	47.5	Image-Event	Voxel Grid	28.2	76.58	96.54	112.1	74.47	95.57
StitchFusion	MiT-B4	64.4	Image-Event	Voxel Grid	36.0	77.41	96.50	143.1	73.69	95.44
OmniSegmentor	DFormer-L	40.9	Image-Event	Voxel Grid	25.3	79.34	96.54	100.7	75.31	95.74
RGB-E Method
ESS	E2VID	52.7	Image-Event	Voxel Grid	20.0	60.43	90.37	77.4	53.29	89.37
EDCNet	ResNet-18	23.1	Image-Event	Voxel Grid	21.3	61.99	93.80	72.5	56.75	92.39
HALSIE	Customed	5.0	Image-Event	Voxel Grid	4.9	60.66	92.50	37.7	52.43	89.01
SE-Adapter	ViT-B	88.1	Image-Event	MSP	62.3	69.06	95.32	267.4	69.77	93.58
EISNet	MiT-B2	34.4	Image-Event	AET	16.8	75.03	96.04	67.3	73.07	95.12
EISNet	MiT-B3	95.0	Image-Event	AET	29.3	76.14	96.14	116.6	73.33	95.07
EISNet	MiT-B4	128.5	Image-Event	AET	37.0	76.06	95.89	147.2	72.56	95.05
Hybrid-Seg	ERViT-T	2.1	Image-Event	Voxel Grid	4.0	67.31	95.07	14.4	66.57	94.27
ESC	MiT-B2	56.9	Image-Event	Voxel Grid	18.3	76.81	96.35	95.1	73.55	95.49
MambaSeg	VMamba-T	25.4	Image-Event	Voxel Grid	33.8	77.56	96.33	159.5	75.10	95.71
Evita-Pico	Ours-P	1.3	Image-Event	Voxel Grid	0.6	71.26	95.78	2.6	65.55	94.12
Evita-Nano	Ours-N	4.9	Image-Event	Voxel Grid	2.3	74.84	96.17	9.2	70.09	94.99
Evita-Tiny	Ours-T	6.5	Image-Event	Voxel Grid	3.2	77.04	96.45	12.9	73.90	95.55
Evita-Small	Ours-S	21.9	Image-Event	Voxel Grid	10.3	78.56	96.72	40.9	75.07	95.70
Evita-Base	Ours-B	34.3	Image-Event	Voxel Grid	14.7	79.11	96.73	58.5	76.08	95.86
Evita-Large	Ours-L	44.3	Image-Event	Voxel Grid	20.8	80.12	96.68	82.7	76.80	95.97

BibTeX

@misc{peng2026evita,
  title  = {Weaving Light and Time: Unified Harmonic-Geometric Representation Learning for Dense RGB-Event Parsing},
  author = {Peng, Chenxu and Zhou, Chongtian and Liu, Dicheng and Yin, Bowen and Dai, Yimian and Liu, Xialei and Cheng, Ming-Ming and Li, Xiang},
  year   = {2026}
}