Result of AcfNet

Model Info

• Note: Test on GTX1080Ti, with resolution 384x1248.

Model Name	FLOPS	Parameters	FPS	Time(ms)
AcfNet(uniform)	1080.0G	5.227M	1.66	600.8
AcfNet(adaptive)	1239.0G	5.559M	1.38	723.1

Experiment

hints

• batch size: n * m, n GPUs m batch/GPU
• pass: clean pass or final pass of Scene Flow dataset, default clean pass
• weight init: initialize the convolution/bn layer while training from scratch, default no initialization
• synced bn: weather use synced bn provided by apex, default False
• float16: weather use mixture precision training with level 01 provided by apex, default False
• scale loss: the loss scale factor when using apex
• time: time consuming including the training and evaluating time, in format: x h(hour) y m(minute)
• EPE: end-point-error for SceneFlow
• xPE (1PE, 2PE, 3PE, 5PE): pixel error where (GroundTruth - Estimation) <= x
• D1(all): 3PE(px) & 5% for KITTI 2015

SceneFlow

10 epoch

RMSProp, lr(10 epochs) schedule: 1-10 with lr*1

model name	lr	batch size	weight init	synced bn	float16	loss scale	EPE(px)	time	BaiDuYun	GoogleDrive
adaptive	0.001	4*3	✗	✓	✗	✗	0.8308	68h18m	link, pw: qxxr	link
uniform	0.001	4*3	✗	✓	✗	✗	0.8511	26h50m	link, pw: 9s4e	link

Disparity Predictor Ablation

If we alternate the disparity predictor from FasterSoftArgmin to LocalSoftArgmin only for reference

model name	predictor	1PE	2PE	3PE	5PE	EPE(px)
adaptive	Faster	7.905	5.125	3.991	2.873	0.8308
adaptive	Local
uniform	Faster	8.626	5.544	4.291	3.061	0.8511
uniform	Local	5.983	3.620	2.838	2.164	0.8216

Analysis

1. Little difference for EPE, but make significant effect on xPE.

2. Therefore, to get better result on KITTI, alternate the disparity predictor from FasterSoftArgmin to LocalSoftArgmin

3. LocalSoftArgmin only works when cost volume supervised with uni-modal distribution, worse result for PSMNet

20 epoch

RMSProp, lr(20 epochs) schedule: 1-20 with lr*1

model name	lr	batch size	weight init	synced bn	float16	loss scale	EPE(px)	time	BaiDuYun	GoogleDrive
adaptive	0.001	4*3	✗	✓	✗	✗	0.7172	134h31m	link, pw: qxxr	link
uniform	0.001	4*3	✗	✓	✗	✗	0.7440	56h53m	link, pw: 9s4e	link

Disparity Predictor Ablation

If we alternate the disparity predictor from FasterSoftArgmin to LocalSoftArgmin only for reference

model name	predictor	1PE	2PE	3PE	5PE	EPE(px)
adaptive	Faster	6.918	4.480	3.485	2.498	0.7172
adaptive	Local
uniform	Faster	7.647	4.917	4.381	2.693	0.7440
uniform	Local	5.338	3.232	2.536	1.927	0.7161

KITTI-2015

model name	lr	batch size	weight init	synced bn	float16	loss scale	D1(all)	time	BaiDuYun	GoogleDrive
adaptive	0.001	4*3	✗	✓	✗	✗
uniform	0.001	4*3	✗	✓	✗	✗

How TO

Alternate the disparity predictor from FasterSoftArgmin to LocalSoftArgmin

In config file, change

disp_predictor=dict(
        # default FasterSoftArgmin
        type="FASTER",
        # the maximum disparity of disparity search range
        max_disp=max_disp,
        # the start disparity of disparity search range
        start_disp=0,
        # the step between near disparity sample
        dilation=1,
        # the temperature coefficient of soft argmin
        alpha=1.0,
        # whether normalize the estimated cost volume
        normalize=True,
    ),

to

disp_predictor=dict(
        # LocalSoftArgmin
        type="LOCAL",
        # the maximum disparity of disparity search range
        max_disp=max_disp,
        # the radius of window when local sampling
        radius=3,
        # the start disparity of disparity search range
        start_disp=0,
        # the step between near disparity sample
        dilation=1,
        # the step between near disparity index when local sampling
        radius_dilation=1,
        # the temperature coefficient of soft argmin
        alpha=1.0,
        # whether normalize the estimated cost volume
        normalize=True,
    ),