LOADING

282 words

1 minute

LOSP (Learnable Orthogonal Subspace Projection)

2026-02-04

No Tags

核心思想#

高光是加法项: I = R⊙L + S，传统log方法无法分离
假设高光向量S在RGB空间具有低秩特性（方向相对固定）
学习投影基W，使得 W·S ≈ 0，物理消除高光
投影后的特征可继续使用YOLA的Log-Ratio逻辑

模块组成#

LOSPProjection: 可学习正交投影层（核心创新）
DualInvariantExtractor: 双流特征提取（光谱+空间不变量）
AdaptiveFusion: 自适应融合
SoftTailLoss: 可微的长尾抑制损失
OrthogonalityLoss: 正交多样性约束

数学基础#

投影: proj_W(I) = W·I，其中W是单位正交基
光谱不变量: log(C_i) - log(C_j) 消除乘法光照
空间不变量: ∇(log(C)) 保留单通道纹理

指标#

baseline的baseline#

yolov3#

类别	GT数量	检测数量	Recall	AP
Bicycle	418	1311	0.866	0.801
Boat	515	1221	0.817	0.722
Bottle	433	1382	0.813	0.724
Bus	164	511	0.909	0.848
Car	919	2429	0.862	0.783
Cat	425	1348	0.831	0.669
Chair	609	2970	0.772	0.651
Cup	356	1084	0.831	0.711
Dog	490	1296	0.898	0.786
Motorbike	242	1435	0.826	0.641
People	2235	6152	0.858	0.768
Table	311	2079	0.746	0.494
mRecall			0.836
mAP				0.717

yolov3纯IIM#

类别	GT数量	检测数量	Recall	AP
Bicycle	418	5830	0.713	0.491
Boat	515	11495	0.713	0.338
Bottle	433	9673	0.637	0.393
Bus	164	3939	0.811	0.586
Car	919	22777	0.691	0.394
Cat	425	10333	0.706	0.165
Chair	609	18323	0.696	0.368
Cup	356	14124	0.643	0.315
Dog	490	10147	0.810	0.349
Motorbike	242	13163	0.624	0.175
People	2235	33656	0.723	0.419
Table	311	12463	0.640	0.180
mRecall			0.701
mAP				0.348

TOOD#

类别	GT数量	检测数量	Recall	AP
Bicycle	418	3331	0.911	0.802
Boat	515	6075	0.926	0.765
Bottle	433	5791	0.868	0.720
Bus	164	2208	0.970	0.892
Car	919	9996	0.930	0.780
Cat	425	3840	0.894	0.725
Chair	609	11617	0.890	0.672
Cup	356	5500	0.904	0.713
Dog	490	4477	0.971	0.861
Motorbike	242	5284	0.917	0.624
People	2235	22387	0.906	0.762
Table	311	10395	0.897	0.501
mRecall			0.915
mAP				0.735

YOLA-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

YOLA-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

LOSP-TOOD初版-未融合原始RGB#

类别	GT数量	检测数量	Recall	AP
Bicycle	418	4115	0.933	0.819
Boat	515	7038	0.951	0.789
Bottle	433	6269	0.882	0.741
Bus	164	2061	0.976	0.896
Car	919	9976	0.939	0.800
Cat	425	5306	0.911	0.724
Chair	609	12550	0.910	0.690
Cup	356	6336	0.919	0.709
Dog	490	4977	0.969	0.858
Motorbike	242	5505	0.930	0.651
People	2235	24726	0.923	0.782
Table	311	9962	0.900	0.512
mRecall			0.929
mAP				0.748

LOSP-yolov3初版-未融合原始RGB#

类别	GT数量	检测数量	Recall	AP
Bicycle	418	1271	0.888	0.816
Boat	515	1361	0.845	0.740
Bottle	433	1635	0.813	0.714
Bus	164	480	0.909	0.859
Car	919	2508	0.887	0.801
Cat	425	1334	0.849	0.685
Chair	609	3055	0.801	0.677
Cup	356	1373	0.846	0.698
Dog	490	1545	0.914	0.784
Motorbike	242	1412	0.781	0.608
People	2235	6582	0.864	0.774
Table	311	2395	0.765	0.448
mRecall			0.847
mAP				0.717

LOSP-TOODv2#

并未改善

LOSP-yolov3v2#

并未改善

LOSP (Learnable Orthogonal Subspace Projection)

/blog/posts/科研笔记/losp/

Author

Zenfish

Published at

2026-02-04

License

CC BY-NC-SA 4.0

Some information may be outdated

VSCode SSH 远程连接与免密登录配置

鞭笞AI给我思路