Last Layer Optical Flow

Task

Under this setting, we only evaluate methods' optical flow predictions for LayeredFlow's last layer points, which are points associated with non-transparent materials.

Metrics

Average end-point-error (EPE) measures the average L2 distance between predicted and ground truth optical flow.
Bad-τ represents the percentage of pixels having L2 error larger than a threshold of τ. We use τ=1, 3, and 5 pixels.

Leaderboard

The categories Behind Transparent contains last layer points that are behind at least one transparent layer.

	All				Reflective				Diffuse				Behind Transparent
Method	EPE↓	1px↓	3px↓	5px↓	EPE↓	1px↓	3px↓	5px↓	EPE↓	1px↓	3px↓	5px↓	EPE↓	1px↓	3px↓	5px↓
FlowNet-C ^[1]	14.94	94.87	75.44	61.66	13.54	94.89	78.26	65.62	17.13	94.82	68.83	52.36	16.37	93.95	64.6	43.73
RAFT ^[2]	8.28	58.68	34.96	28.86	8.25	61.68	39.38	32.18	9.58	51.66	24.57	21.06	8.48	52.19	19.54	15.59
FlowFormer ^[3]	7.96	59.65	35.4	27.64	9.28	62.17	39.66	31.5	6.26	53.74	25.4	18.57	8.12	54.14	23.24	14.94
GMFlow+ ^[4]	8.52	86.11	57.6	42.52	9.46	88.76	62.08	45.13	7.64	79.89	47.09	36.4	5.81	79.23	45.19	30.77
GMFlow ^[5]	7.69	82.16	52.94	39.09	8.52	85.53	57.77	42.16	7.11	74.24	41.6	31.89	6.15	74.07	39.32	26.02
CRAFT ^[6]	9.8	62.43	36.66	30.27	9.82	64.59	40.21	33.22	7.66	57.36	28.32	23.36	6.87	54.48	24.14	16.7
SKFlow ^[7]	8.89	59.83	34.78	29.29	9.14	62.37	38.13	31.93	7.2	53.86	26.92	23.1	7.38	53.8	19.88	16.11
GMA ^[8]	7.55	62.0	35.99	29.13	7.95	65.12	40.63	32.77	9.51	54.67	25.11	20.56	6.81	54.81	20.09	13.8
PWC-Net ^[9]	19.39	72.59	48.44	40.1	14.98	74.06	50.42	42.2	23.78	69.15	43.79	35.11	24.87	68.12	38.06	29.1
FlowNet2 ^[10]	14.24	80.88	58.87	48.59	13.21	84.79	63.42	52.45	14.98	71.7	48.18	39.54	13.67	70.28	44.35	28.4
RAFT-ft. (L) ^[11]	7.72	57.67	34.32	27.15	9.43	59.71	37.09	29.2	7.79	52.88	27.8	22.33	4.34	52.07	19.01	12.9

[1] Flownet: Learning optical flow with convolutional networks. [paper]

[2] Raft: Recurrent all-pairs field transforms for optical flow. [paper] [code]

[3] FlowFormer: A Transformer Architecture for Optical Flow. [paper] [code]

[4] Unifying flow, stereo and depth estimation. [paper] [code]

[5] Gmflow: Learning optical flow via global matching. [paper] [code]

[6] Craft: Cross-attentional flow transformer for robust optical flow. [paper] [code]

[7] Skflow: Learning optical flow with super kernels.. [paper] [code]

[8] Learning to estimate hidden motions with global motion aggregation. [paper] [code]

[9] Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. [paper] [code]

[10] Flownet 2.0: Evolution of optical flow estimation with deep networks. [paper]

[11] LayeredFlow: A Real-World Benchmark for Non-Lambertian Multi-Layer Optical Flow. [paper] [code]