Active Reinforcement Learning for the Semantic Segmentation of Urban Images

Figures & data

Figure 1. A street view (left) and its semantic segmentation map (right) (Cityscapes dataset).

Figure 2. Imbalanced class distribution in the Cityscapes dataset.

Figure 3. Scene examples from the Cityspaces dataset.

Figure 4. Class distribution in the Cityscapes training set.

Figure 5. A scene (left) and its semantic segmentation map (right) (GTAv dataset).

Figure 6. The architecture of end-to-end Active Reinforcement Learning model.

Figure 7. Query network architecture.

Figure 8. Intersection over union.

Figure 9. Implementation flowchart.

Table 1. The Cityscapes dataset partition.

DS	DT	DR	DV
10	150	200	2615

Table 2. Pretraining details.

	Backbone	Input Size	Learning Rate	Loss	Batch size	Optimizer	Validation Mean IoU
FPN	ResNet-50	1904 $\times$1024	10 $\times$ e-3,10 $\times$ e-4	Weighted cross entropy	4	Adam	73%
DeepLabV3	ResNet-50	1904 $\times$ 1024	10 $\times$ e-3,10 $\times$ e-4	Weighted cross entropy	4	Adam	73%
DeepLabV3+	ResNet-50	1904 $\times$ 1024	10 $\times$ e-3,10 $\times$ e-4	Weighted cross entropy	4	Adam	74%

Figure 10. Pre-training FPN, DeepLabV3, and DeepLabV3+ segmentation networks on GTAv synthetic dataset.

Figure 11. Fine-tuning on D_t subset.

Table 3. Fine-tuning details.

	Backbone	Input Size	Learning Rate	Loss	Batch size	Optimizer	Validation Mean IoU
FPN	ResNet-50	512 $\times$1024	10 $\times$ e-5, 10 $\times$ e-6	Weighted cross entropy	16	Adam	52.2%
DeepLabV3	ResNet-50	512 $\times$ 1024	10 $\times$ e-5, 10 $\times$ e-6	Weighted cross entropy	16	Adam	54.4%
DeepLabV3+	ResNet-50	512 $\times$ 1024	10 $\times$ e-5, 10 $\times$ e-6	Weighted cross entropy	16	Adam	64.6%

Table 4. Percentage of region numbers from all of the dataset.

Regions number	1920	3840	7680	11520	19200	30720
Percentage	0.5%	1%	2%	3%	5%	8%

Table 5. Comparison of performance of the FPN network across various methods with different budgets (MIoU).

Budget	Supervised	Random	Entropy	BALD	MioU RALis	FWA IoU RALis
1920	0.28	0.53	0.59	0.56	0.58	0.60
3840	0.29	0.57	0.60	0.58	0.61	0.62
7680	0.30	0.61	0.64	0.62	0.65	0.65
11520	0.26	0.62	0.65	0.64	0.66	0.65
19200	0.30	0.65	0.67	0.55	0.67	0.68
30720	0.28	0.64	0.66	0.66	0.67	0.71

Figure 12. Performance of our method with the FPN segmentation network compared to baselines.

Figure 13. The performance of the RALis method rewarded by MIoU and FWA IoU.

Table 6. Performance of our method on six underrepresented classes (MIoU).

Budget	Bicycle	Motorcycle	Train	Bus	Truck	Rider	Total
1920	0.39	0.34	0.28	0.50	0.38	0.61	0.42
3840	0.40	0.37	0.32	0.54	0.48	0.64	0.46
7680	0.42	0.46	0.41	0.63	0.52	0.66	0.52
11520	0.44	0.54	0.44	0.46	0.65	0.48	0.66
19200	0.54	0.51	0.49	0.72	0.59	0.69	0.59
30720	0.54	0.63	0.61	0.72	0.52	0.72	0.62
Pixel %	0.41	0.10	0.23	0.24	0.27	0.14	1.39

Figure 14. Our RALis method with DeepLabV3 segmentation network in comparison to other baselines.

Table 7. The comparison of performance of the DeepLabV3 network across various methods with different budgets (MIoU).

Budget	Supervised	Random	Entropy	BALD	FWA IoU RALis
1920	0.29	0.55	0.54	0.30	0.59
3840	0.28	0.54	0.57	0.58	0.62
7680	0.29	0.53	0.60	0.61	0.65
11520	0.28	0.59	0.59	0.62	0.65
19200	0.29	0.58	0.63	0.63	0.65
30720	0.30	0.61	0.63	0.65	0.67

Figure 15. RALis method with a DeepLabV3+ segmentation network in comparison to other baselines.

Table 8. The comparison of performance of the DeepLabV3 network across various methods with different budgets (MIoU).

Budget	Supervised	Random	Entropy	BALD	FWA IoU RALis
1920	0.23	0.58	0.55	0.56	0.58
3840	0.27	0.61	0.55	0.55	0.56
7680	0.27	0.60	0.55	0.58	0.59
11520	0.28	0.62	0.61	0.61	0.65
19200	0.31	0.63	0.63	0.66	0.69
30720	0.31	0.65	0.65	0.66	0.72

Figure 16. The segmentation models are trained on all of the Cityscapes dataset, containing 2975 images. The results are compared to IoU for validation.

Figure 17. A comparison of the performance of the RALis method rewarded with weighted IoU.

Table 9. The performance of supervised and FWA RALis methods across three networks (MIoU).

	Supervised (100% datd)	FWA RALis (8% data)
FPN	71.8%	70.9%
DeepLabV3	73.2%	67.3%
DeepLabV3+	75.1%	71.8%

Table 10. The performance of baselines and RALis method across three networks with 8% budget of all data (MIoU).

	Random	Entropy	BALD	FWA IoU RALis
FPN	64%	66%	66%	71%
DeepLabV3	61%	63%	65%	67%
DeepLabV3+	65%	65%	66%	72%