指标

FRBNet

Paradigm	Method	ExDark(YOLOv3)Recall	ExDark(YOLOv3)mAP	ExDark(TOOD)Recall	ExDark(TOOD)mAP	DarkFace(YOLOv3)Recall	DarkFace(YOLOv3)mAP	DarkFace(TOOD)Recall	DarkFace(TOOD)mAP
	Baseline	84.6	71.0	91.9	72.5	73.8	54.8	80.9	57.0
Enhancement	SMG[66] (CVPR-23)	82.3	68.5	91.8	71.5	73.4	52.4	80.2	56.3
	NeRCo[70] (ICCV-23)	83.4	68.5	91.8	71.8	73.8	53.0	79.4	56.8
	LightDiff[25] (ECCV-24)	84.3	71.3	92.1	72.9	75.5	57.4	81.0	58.7
	DarkIR[15] (CVPR-25)	81.9	68.2	90.9	72.0	74.5	55.9	81.4	60.4
Synthetic Data	DAINet[13] (CVPR-24)*	86.7	73.4	-	-	74.8	56.9	-	-
	WARLearn[1] (WACV-25)	85.6	72.4	92.8	73.4	74.5	56.2	80.8	59.4
Multi-task	MAET[10] (ICCV-21)	85.1	72.5	92.5	74.3	74.7	55.7	80.7	59.6
	IAT[9] (BMVC-22)	85.0	72.6	92.9	73.0	73.6	55.5	79.7	58.3
Plug-and-play	DENet[46] (ACCV-22)	84.2	71.3	92.6	73.5	71.8	52.6	73.6	49.6
	FeatEnHancer[21] (ICCV-23)	90.4	71.2	96.4	74.6	74.1	55.2	81.7	60.5
	YOLA[5] (NeurIPS-24)	86.1	72.7	93.8	75.2	74.9	56.3	83.1	63.2
	FRBNet(NeurIPS-25)	90.6	74.9	93.2	75.4	75.7	57.7	82.7	65.1

上图是FRBNet中的指标

下面我跑出来的测试指标，比论文图降低了0.2个百分点

yolov3-只开源了yolov3代码，tood怕补充错，不补充

class	gts	dets	recall	ap
Bicycle	418	1353	0.902	0.834
Boat	515	1607	0.874	0.756
Bottle	433	2021	0.850	0.750
Bus	164	572	0.951	0.886
Car	919	2845	0.922	0.827
Cat	425	1150	0.845	0.695
Chair	609	3678	0.857	0.711
Cup	356	1697	0.860	0.724
Dog	490	1498	0.931	0.832
Motorbike	242	1351	0.822	0.669
People	2235	7391	0.894	0.804
Table	311	2926	0.823	0.483
mAP				0.747

YOLA

Exdark-iou_thr:0.5

TOOD

类别	GT数量	检测数量	Recall	AP
Bicycle	418	3584	0.933	0.833
Boat	515	5438	0.940	0.785
Bottle	433	5793	0.891	0.742
Bus	164	2042	0.957	0.888
Car	919	9011	0.933	0.794
Cat	425	3733	0.908	0.749
Chair	609	10100	0.893	0.706
Cup	356	5269	0.899	0.722
Dog	490	4176	0.963	0.862
Motorbike	242	5374	0.905	0.633
People	2235	20178	0.905	0.778
Table	311	8944	0.884	0.495
mRecall			0.918
mAP				0.749↓

yolov3

类别	GT数量	检测数量	Recall	AP
Bicycle	418	1159	0.895	0.832
Boat	515	1275	0.860	0.770
Bottle	433	1404	0.818	0.726
Bus	164	446	0.939	0.881
Car	919	2327	0.890	0.816
Cat	425	1407	0.845	0.685
Chair	609	2648	0.788	0.668
Cup	356	1201	0.820	0.699
Dog	490	1345	0.906	0.801
Motorbike	242	1182	0.798	0.638
People	2235	6357	0.868	0.778
Table	311	2123	0.807	0.494
mRecall			0.853
mAP				0.732↑

降点（得靠大batchsize来增点，点数也没之前高，离谱，之前不知道改啥了）

DarkFace-iou_thr:0.5

TOOD（逆天成绩，之前没跑出来过，莫名其妙）

类别	GT数量	检测数量	Recall	AP
Face	4205	49304	0.868	0.711
mRecall			0.868
mAP				0.711

yolov3

class	gts	dets	recall	ap
Face	4205	22115	0.788	0.605
mAP				0.605↓

OURS

Exdark-iou_thr:0.5

TOOD

类别	GT数量	检测数量	Recall	AP
Bicycle	418	3397	0.933	0.843
Boat	515	5705	0.932	0.797
Bottle	433	5247	0.880	0.739
Bus	164	2094	0.957	0.887
Car	919	8203	0.938	0.801
Cat	425	3743	0.882	0.731
Chair	609	10632	0.892	0.698
Cup	356	5009	0.907	0.708
Dog	490	3953	0.973	0.880
Motorbike	242	4965	0.905	0.645
People	2235	21588	0.912	0.779
Table	311	8723	0.904	0.509
mRecall			0.918
mAP				0.751↑

yolov3

类别	GT数量	检测数量	Recall	AP
Bicycle	418	1160	0.892	0.829
Boat	515	1264	0.835	0.752
Bottle	433	1483	0.804	0.706
Bus	164	459	0.933	0.878
Car	919	2169	0.884	0.799
Cat	425	1367	0.842	0.678
Chair	609	2745	0.782	0.660
Cup	356	1175	0.840	0.703
Dog	490	1403	0.902	0.794
Motorbike	242	1123	0.835	0.672
People	2235	6184	0.872	0.783
Table	311	2147	0.775	0.492
mRecall			0.850
mAP				0.729↓

DarkFace-iou_thr:0.5

TOOD

类别	GT数量	检测数量	Recall	AP
Face	4205	49120	0.870	0.715
mRecall			0.870
mAP				0.715↑

yolov3

类别	GT数量	检测数量	Recall	AP
Face	4205	21860	0.796	0.622
mRecall			0.796
mAP				0.622

Ablation

TOOD

yola_hfe_exdark_ablation1_reflected_only

类别	GT数量	检测数量	Recall	AP
Bicycle	418	3330	0.935	0.837
Boat	515	5060	0.936	0.788
Bottle	433	5127	0.880	0.732
Bus	164	2179	0.976	0.892
Car	919	9108	0.936	0.795
Cat	425	3613	0.904	0.731
Chair	609	9965	0.883	0.700
Cup	356	5064	0.902	0.705
Dog	490	4143	0.963	0.871
Motorbike	242	5237	0.926	0.660
People	2235	20509	0.908	0.782
Table	311	9362	0.897	0.506
mRecall			0.920
mAP				0.750↑

work_dirs/yola_hfe_exdark_ablation2_plus_frequency

类别	GT数量	检测数量	Recall	AP
Bicycle	418	3408	0.933	0.833
Boat	515	5927	0.940	0.784
Bottle	433	5877	0.871	0.729
Bus	164	2023	0.963	0.879
Car	919	8669	0.936	0.796
Cat	425	4003	0.892	0.748
Chair	609	10884	0.901	0.708
Cup	356	5210	0.910	0.718
Dog	490	4219	0.967	0.876
Motorbike	242	4970	0.905	0.627
People	2235	21489	0.908	0.778
Table	311	8564	0.897	0.511
mRecall			0.919
mAP				0.749=

yola_hfe_exdark_ablation3_plus_wavelet

类别	GT数量	检测数量	Recall	AP
Bicycle	418	3475	0.940	0.835
Boat	515	6092	0.938	0.785
Bottle	433	5430	0.887	0.738
Bus	164	2100	0.951	0.887
Car	919	8787	0.930	0.793
Cat	425	3892	0.896	0.747
Chair	609	10453	0.901	0.706
Cup	356	5225	0.904	0.715
Dog	490	4098	0.969	0.880
Motorbike	242	4978	0.921	0.644
People	2235	20963	0.906	0.778
Table	311	8572	0.900	0.512
mRecall			0.920
mAP				0.752↑

yola_hfe_exdark_ablation4_three_branches_no_attn

类别	GT数量	检测数量	Recall	AP
Bicycle	418	3255	0.928	0.831
Boat	515	5958	0.930	0.784
Bottle	433	5981	0.882	0.749
Bus	164	2169	0.970	0.890
Car	919	8849	0.937	0.805
Cat	425	3797	0.906	0.747
Chair	609	10142	0.901	0.699
Cup	356	5254	0.904	0.731
Dog	490	4136	0.971	0.876
Motorbike	242	5134	0.921	0.634
People	2235	21825	0.914	0.782
Table	311	8585	0.900	0.492
mRecall			0.922
mAP				0.752↑

yola_hfe_exdark_ablation5_reflected_with_attn

类别	GT数量	检测数量	Recall	AP
Bicycle	418	3339	0.931	0.837
Boat	515	5855	0.942	0.779
Bottle	433	5389	0.873	0.741
Bus	164	2110	0.951	0.892
Car	919	9019	0.941	0.793
Cat	425	3747	0.904	0.759
Chair	609	10270	0.895	0.706
Cup	356	5178	0.902	0.714
Dog	490	3961	0.965	0.872
Motorbike	242	4855	0.926	0.646
People	2235	20443	0.910	0.779
Table	311	8296	0.897	0.501
mRecall			0.920
mAP				0.752↑

新思路

路由融合

只使用yola原始的卷积融合融合多分支特征不能很好的根据样本做出分支特征的取舍，因此加入MoE的路由机制

使用mlp路由

把两个分支的特征分别池化降采样然后展平成48维输入给mlp，输出为二维向量，分别对应分支权重，下面是yolov3的表现，和yola比降低了0.1个百分点

类别	GT数量	检测数量	Recall	AP
Bicycle	418	1123	0.880	0.824
Boat	515	1234	0.841	0.746
Bottle	433	1499	0.827	0.734
Bus	164	443	0.933	0.881
Car	919	2369	0.885	0.813
Cat	425	1185	0.828	0.681
Chair	609	2585	0.796	0.686
Cup	356	1217	0.840	0.714
Dog	490	1306	0.914	0.809
Motorbike	242	1104	0.793	0.637
People	2235	6274	0.867	0.782
Table	311	2126	0.772	0.467
mRecall			0.848
mAP				0.731↓

卷积做空间上的路由

上面感觉 $B * 24 * 608 * 608$ 到 $B * 48$ 作为路由输入信息损失太多，想着用卷积做 $B * b r a n c h n u m s * 608 * 608$ 的路由，而且感觉只利用分支的特征信息做路由感觉不够有依据，按道理还要看原材料是怎么样的然后决定选择什么特征，因此就融入了 $R G B : B * 24 * 608 * 608$ 的特征作为路由输入。

# 3x3 卷积生成空间权重图
# 输入: (B, 48/72, H, W) → 输出: (B, 2, H, W)
self.router_conv = nn.Conv2d(
    actual_in_channels,
    num_experts,
    kernel_size=kernel_size,
    padding=padding,
    bias=True
)

还是降点了，不多

类别	GT数量	检测数量	Recall	AP
Bicycle	418	1135	0.885	0.822
Boat	515	1196	0.839	0.748
Bottle	433	1608	0.801	0.709
Bus	164	434	0.927	0.876
Car	919	2149	0.882	0.808
Cat	425	1300	0.833	0.676
Chair	609	2589	0.782	0.659
Cup	356	1194	0.860	0.722
Dog	490	1357	0.914	0.783
Motorbike	242	1070	0.806	0.647
People	2235	6482	0.864	0.771
Table	311	2070	0.746	0.474
mRecall			0.845
mAP				0.725↓