dk-moku/deep-unsupervised-saliency-detection

[WIP] Implmentation of the Deep Unsupervised Saliency Detection: A Multiple Noisy Labeling Perspective

Requirements

1. pytorch 1.0
2. python 3.6
3. numpy Dataset

---

1. MSRA-B

Results

Table of Contents

1. Training on Small Dataset(10 Images)
2. Training on Large Dataset(1500 Images Training, 500 Images Validataion, 500 Images Test)

Training on Small Dataset(10 Images)

Experiment Performed for overfitting, checking if the model works, tested for all models, Only reporting for the full(Training Noise module)

Training on Large Dataset(1500 Images Training, 500 Images Validataion, 500 Images Test)

The most important part here is the scheduler, since we keep on training with the same leanring rate without, Image size is taken to be 256, Number of Epochs is 20, Early Stopping on validation loss, paitence of 5 epochs

1. Exp-1 Using Adam Optimizer

   Exp Name	Optimizer	Batch Size	LR	Betas	Momentum	Scheduler	Notes	Decay Factor	Paitence	THRESHOLD	MINLR	COOLDOWN	Recall-Test	Precision-Test	F1-Test	MAE-TEST
   Real	Adam	16	3e-4	(0.9, 0.99)	x	ReduceLROnPlateu	Label is the ground truth	0.9	1	1e-4	1e-16	1	0.5023	0.9622	0.66	0.035
   Noise	Adam	8	3e-4	(0.9, 0.99)	x	ReduceLROnPlateu	Use all Noise labesl	0.9	1	1e-4	1e-16	1	0.793	0.946	0.862	0.028
   Avg	Adam	16	3e-4	(0.9, 0.99)	x	ReduceLROnPlateu	Use Avg of Noise Labels	0.9	1	1e-4	1e-16	1	0.802	0.907	0.851	0.041
   Full	Adam	4	3e-4	(0.9, 0.99)	x	ReduceLROnPlateu	Full Training	0.9	1	1e-4	1e-16	1	0.841	0.857	0.848	0.036

Sample Images Average Exp Sample Maps Average Exp Sample [email protected] Real Exp Sample Maps Real Exp Sample Maps [email protected] Noise Exp Sample Maps Noise Exp Sample Maps [email protected] Full Exp Sample Maps Full Exp Sample Maps [email protected]

1. Exp-2 Using SGD Optimizer

   Exp Name	Optimizer	Batch Size	LR	Betas	Momentum	Scheduler	Notes	Decay Factor	Paitence	THRESHOLD	MINLR	COOLDOWN	Recall-Test	Precision-Test	F1-Test	MAE-TEST
   Real	SGD	16	1e-3	x	0.9	ReduceLROnPlateu	Label is the ground truth	0.9	1	1e-4	1e-16	1	0.865	0.907	0.885	0.032
   Noise	SGD	8	1e-3	x	0.9	ReduceLROnPlateu	Use all Noise labesl	0.9	1	1e-4	1e-16	1	0.620	0.820	0.706	0.044
   Avg	SGD	16	1e-3	x	0.9	ReduceLROnPlateu	Use Avg of Noise Labels	0.9	1	1e-4	1e-16	1	0.934	0.639	0.758	0.058
   Full	SGD	4	1e-3	x	0.9	ReduceLROnPlateu	Full Training	0.9	1	1e-4	1e-16	1	0.825	0.795	0.809	0.027

Average Exp Sample Maps Average Exp Sample [email protected] Real Exp Sample Maps Real Exp Sample Maps [email protected] Noise Exp Sample Maps Noise Exp Sample Maps [email protected] Full Exp Sample Maps Full Exp Sample Maps [email protected]