First Layer Optical Flow

Task

Under this setting, we only evaluate methods' optical flow predictions for LayeredFlow's first layer points. For transparent objects, methods must predict the optical flow of the transparent occluder rather than the background, similar to setting of other non-Lambertian benchmarks.

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

	All				Transparent				Reflective				Diffuse
Method	EPE↓	1px↓	3px↓	5px↓	EPE↓	1px↓	3px↓	5px↓	EPE↓	1px↓	3px↓	5px↓	EPE↓	1px↓	3px↓	5px↓
RAFT ^[1]	16.49	78.45	55.64	45.78	20.72	82.72	59.69	49.06	8.51	62.05	40.57	27.56	10.76	50.78	27.56	24.39
FlowNet-C ^[2]	21.14	94.88	77.86	65.2	24.01	94.84	77.9	65.15	13.85	94.82	79.18	67.35	17.04	96.18	70.95	56.46
FlowFormer ^[3]	18.49	78.83	58.61	49.24	22.56	83.02	63.42	53.64	9.54	61.73	39.63	32.24	5.01	56.57	30.21	21.33
GMFlow+ ^[4]	17.62	89.83	67.21	54.29	21.36	90.36	68.83	56.45	9.68	88.65	61.91	45.8	10.06	82.53	52.35	41.21
GMFlow ^[5]	16.92	88.45	64.0	51.71	20.72	89.51	65.9	54.01	8.74	85.86	58.01	43.18	8.29	74.63	45.58	35.64
CRAFT ^[6]	17.82	80.31	57.6	47.9	21.57	84.07	61.86	51.34	10.11	64.79	40.48	33.91	8.73	60.94	33.78	29.49
SKFlow ^[7]	18.14	79.12	57.47	48.33	22.17	83.31	62.01	52.12	9.41	62.38	38.95	32.89	8.17	55.15	33.09	28.21
GMA ^[8]	16.58	79.26	57.04	46.6	20.35	82.93	61.16	49.91	8.18	65.34	41.04	33.83	12.0	55.04	31.45	25.48
PWC-Net ^[9]	28.39	83.93	63.66	54.33	31.75	86.34	66.69	57.1	15.45	74.12	51.02	43.56	20.48	70.8	48.74	37.06
FlowNet2 ^[10]	20.67	86.42	66.54	56.66	23.54	87.19	67.61	57.55	13.52	84.57	63.45	53.97	15.42	76.3	54.82	47.54
RAFT-ft. (L) ^[11]	17.46	78.13	53.12	43.33	18.54	82.15	56.06	45.73	17.3	62.6	41.75	33.89	14.69	52.6	34.73	28.87
RAFT-ft. (S) ^[12]	17.94	79.53	59.47	49.69	21.96	82.94	63.15	52.85	8.89	66.41	45.21	37.11	9.07	57.7	36.44	31.34
RAFT-ft. (S+L) ^[13]	15.63	77.81	52.75	42.76	18.39	81.88	56.17	45.4	11.73	61.93	39.48	32.97	6.95	52.75	31.23	24.24

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

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

[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]

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

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