Final Statistics

	A	B	C	D	E	F	G	H	I	J	K	L	M	N	O	P	Q	R	S
1
2		PlotNet Versions
3		Version	Backbone	Neighbouring Info.	Linking Head	beta1	beta2	mAP@0.9
4		v1	R22 + FPN	NIL	4x2	1	0	80.78	(with avg pool)
5		v2	R22 + FPN	NIL	4x2	1	0	83.89	(without avg pool)
6		v3	R22 + FPN	Neighbouring Info + Conv	4x2	1	0	89.49	(CE loss for linking)
7		v4	R22 + FPN	Neighbouring Info + Conv	8x1	1	0	90.93	(MSE loss for linking)
8		v5	R22 + FPN	Neighbouring Info + Conv	4x2	0.1	1	64.93
9		v6	R22 + FPN	Neighbouring Info + Conv	8x1	0.1	1	72.28
10		v7	R22 + FPN	Neighbouring Info + Conv	8x1	0.8	0.2	89.72
11		v8	R22 + FPN	Neighbouring Info + Conv	8x1	0.9	0.25	90.17
12		v9	R22 + FPN	Neighbouring Info + Conv	8x1	0.9	0.1	86.52
13		v10	R22 + FPN	Neighbouring Info + Conv	8x1	0.7	0.15	87.05
14		v11	R22 + FPN	Neighbouring Info + Conv	8x1	0.4	0.6	85.64
15		v12	R22 + FPN	Neighbouring Info + Conv	8x1	1	0	91.94	(No Regression for Background)
16		v13	R22 + FPN	Neighbouring Info + Conv	8x1	0.9	0.25	76.35	(No Regression for Background + More Conv Layers for makeing neighbour volume + SGD with momentum)
17		v14	R22 + FPN	NIL	4x2	0.9	0.25	86.25	added IOU loss + No regression for bg + PlotNet_v2
18		v15	R22 + FPN	NIL	4x2	0.9	0.25	85.68	added log IOU loss + No regression for bg + PlotNet_v2
19		v16	R22 + FPN	Neighbouring Info + Conv	8x1	1	0	86.86	with averagePool + PlotNet_v4
20		v17	R22 + FPN	Neighbouring Info + Conv	8x1	0.9	0.25		(No Regression for Background + PlotNet_v12)
21		v18	R22 + FPN	NIL	4x2	0	1	44.78	1 - IOU loss + No regression for bg + PlotNet_v2 + L2 regulariser
22		v19	R22 + FPN	NIL	4x2	0	1	50.11	log(IOU) loss + L2 regulariser + No regression for bg + PlotNet_v2
23
24
25
26
27
28
29		Version	Backbone	Neighbouring Info.	Linking Head	Smooth L1	log IoU	1 - IoU	Custom Loss	Gamma	Link Loss	without BG regression	L2 regularizer	Weight Decay	Learning Rate	#Epochs	mAP
30		v23	R22 + FPN	NIL	4x2	1	0	0	0	0.85		1	0	1	0.0001	10	81.41	(minsky)
31		v24	R22 + FPN	NIL	4x2	0	1	0	0	0.85		1	0	1	0.00005	10	81.64
32		v25	R22 + FPN	NIL	4x2	0	0	1	0	0.85		1	0	1	0.00005	10	81.09	(minsky)
33		v26	R22 + FPN	NIL	4x2	0	0	0	1	0.85		1	0	1	0.00005	10	81.71	(minsky)
34		v27	R22 + FPN	Neighbouring Info + Conv	8x1	1	0	0	0	0.85	MSE	1	0	1	0.0001	10	85.73	(gauss)
35		v28	R22 + FPN	Neighbouring Info + Conv	8x1	0	1	0	0	0.85	MSE	1	0	1	0.00005	10	80.34	(gauss)
36		v29	R22 + FPN	Neighbouring Info + Conv	8x1	0	0	1	0	0.85	MSE	1	0	1	0.00005	10	80.87	(gauss)
37		v30	R22 + FPN	Neighbouring Info + Conv	8x1	0	0	0	1	0.85	MSE	1	0	1	0.00005	6	83.38	(gauss)	82.20 (ep10)
38		v31	R22 + FPN	Neighbouring Info + Conv	8x1	1	0	0	0.6	0.85	MSE	1	0	1	0.0001	10	93.15
39		v32	R22 + FPN	Neighbouring Info + Conv	8x1	0	0	0	1	2	MSE	1	0	0	0.0001	10	90.88
40		v33	R22 + FPN	Neighbouring Info + Conv	8x1	1	0.6	0	0	-	MSE	1	0	0	0.0001	10	93.29	(Experiment with the contribution of IOU loss, i.e change 0.6 to something else)
41		v34	R22 + FPN	Neighbouring Info + Conv	8x1	0	0	0	1	3	MSE	1	0	0	0.0001	10	87.55
42		v35	R22 + FPN	Neighbouring Info + Conv	8x1	0	0	0	1	4	MSE	1	0	0	0.0001	10	90.14
43		v36	R22 + FPN	Neighbouring Info + Conv	8x1	1	0	0	0.6	2	MSE	1	0	1	0.0001		91.24
44		v37	R22 + FPN	Neighbouring Info + Conv	8x1	0.6	0	0	1	0.85	MSE	1	0	1	0.0001	10	90.28
45		v38	R22 + FPN	Neighbouring Info + Conv	8x1	1	0	0.6	0	-	MSE	1	0	1	0.0001	10	93.06
46		v39	R22 + FPN	NIL	4x2	1	0	0	0.6	0.85		1	0	0	0.0001	10	84.06
47		v40	R22 + FPN	NIL	4x2	1	0.6	0	0	-		1	0	0	0.0001	10	85.41
48		v41	R22 + FPN	NIL	4x2	1	0	0.6	0	-		1	0	0	0.0001	10	86.81
49		v42	R22 + FPN	Neighbouring Info + Conv	8x1	1	1	0	0	-	MSE	1	0	0	0.0001	10	88.44
50		v43	R22 + FPN	Neighbouring Info + Conv	8x1	1	0	0	1	2	MSE	1	0	1	0.0001		93.39
51		v44	R22 + FPN	Neighbouring Info + Conv	8x1	0	0	0	1	1	MSE	1	0	0	0.0001	10
52	Gamma Tuning	v45	R22 + FPN	Neighbouring Info + Conv	8x1	0	0	0	1	1.25	MSE	1	0	0	0.0001	10	84.63
53		v46	R22 + FPN	Neighbouring Info + Conv	8x1	0	0	0	1	1.5	MSE	1	0	0	0.0001	10	84.49
54		v47	R22 + FPN	Neighbouring Info + Conv	8x1	0	0	0	1	1.75	MSE	1	0	0	0.0001	10	90.14
55
56		v48	R22 + FPN	Neighbouring Info + Conv	8x1	1	0	1	0	-	MSE	1	0	1	0.0001	10	92.19
57		v49	R22	Neighbouring Info + Conv	8x1	1	0	0	0	0	MSE	1	0	1	0.0001	10
58		v50	R22 + FPN	Neighbouring Info + Conv	8x1	0	0	0	GIOU	-	MSE	1	0	1	0.0001	10	88.21
59		v51	R22 + FPN	Neighbouring Info + Conv	8x1	1	0	0	GIOU	-	MSE	1	0	1	0.0001	10	93.34
60		v52	R22 + FPN	Neighbouring Info + Conv	8x1	0	0	0	DIOU	-	MSE	1	0	1	0.0001	10
61		v53	R22 + FPN	Neighbouring Info + Conv	8x1	0	0	0	CIOU	-	MSE	1	0	1	0.0001	10
62		v54	R22 + FPN	Neighbouring Info + Conv	8x1	1	0	0	DIOU	-	MSE
63	no_linking_head with L1	v55	R22 + FPN	Neighbouring Info + Conv	8x1	1	0	0	0	-	-	1	0	1	0.0001	10
64		v56	R22 + FPN	Neighbouring Info + Conv	8x1	1	0	0	CIOU	-	MSE	1
65
66		v57	R22 + FPN	Neighbouring Info + Conv	8x1	0	0	0	CustomLossDIOU	-	MSE	1
67		v58	R22 + FPN	Neighbouring Info + Conv	8x1	1	0	0	CustomLossDIOU	-	MSE	1
68		v59	R22 + FPN	Neighbouring Info + Conv	8x1	0	0	0	CustomLossCIOU	-	MSE
69		v60	R22 + FPN	Neighbouring Info + Conv	8x1	1	0	0	CustomLossCIOU	-	MSE	1
70	v55 + no_linking_head with L1+custom	v61	R22 + FPN	Neighbouring Info + Conv	8x1	1	0	0	1	-	-	1	0	1	0.0001	10
71		v62								2.25
72		v63								2.5
73		v64								2.75
74													0.5	0.75	0.9
75												GIOU
76												DIOU
77				IOU@0.5	IOU@0.75	IOU@0.9						CIOU
78			Smooth-L1	98.24	97.74	91.77
79			1-IOU			80.87
80			-log IOU			80.34
81			GIOU			88.21
82			DIOU			89.5
83			CIOU			90
84			Custom	98.31	97.3	90.88
85
86
87								mAP@0.9 (in %)
88						Relative improvement at IOU@0.9		Models	Before Normalization	After Normalization
89					Smooth-L1	-0.97		Smooth-L1 (SL1)	x	x
90					1-IOU	12.38		SL1 + 1-IOU	x	x
91					-log IOU	13.12		SL1 + -log IOU	x	x
92					GIOU	3.03		SL1 + GIOU	x	x
93					DIOU	1.54		SL1 + CIOU	x	x
94					CIOU	0.98		SL1 + DIOU	x	x
95								SL1 + proposed loss	x	x
96
97
98
99
100