The effect of high-resolution images on results

[fy99925]

I used A100 on the server to train the original resolution image (almost 5600*3600), with all parameters unchanged and 30,000 iterations. The result is blurrier than the image training that is automatically cropped to 1.6k. It is strange. I think higher resolution images may lead to better results.

I wonder about this, what is the effect of high resolution images on guassian splatting? How can I use high resolution images better in gs?

[jaco001]

What parameter you use?

[fy99925]

Hi!

I actually used a derivative work of 3DGS and modified it a bit. But I think it's probably a common problem, which may be caused by some mechanism in 3DGS. But I didn't figure it out, and it confused me.

The parameters are basically the default for this job and have not been modified. The parameters are as follows:

class ModelParams(ParamGroup): 
    def __init__(self, parser, sentinel=False):
        self.sh_degree = 3
        self._source_path = ""
        self._model_path = ""
        self._images = "images"
        self._resolution = -1
        self._white_background = False
        self.data_device = "cuda"
        self.eval = False
        self.preload_img = True
        self.ncc_scale = 1.0
        self.multi_view_num = 8
        self.multi_view_max_angle = 30
        self.multi_view_min_dis = 0.01
        self.multi_view_max_dis = 1.5
        super().__init__(parser, "Loading Parameters", sentinel)

    def extract(self, args):
        g = super().extract(args)
        g.source_path = os.path.abspath(g.source_path)
        return g

class PipelineParams(ParamGroup):
    def __init__(self, parser):
        self.convert_SHs_python = False
        self.compute_cov3D_python = False
        self.debug = False
        super().__init__(parser, "Pipeline Parameters")

class OptimizationParams(ParamGroup):
    def __init__(self, parser):
        self.iterations = 30_000
        self.position_lr_init = 0.00016
        self.position_lr_final = 0.0000016
        self.position_lr_delay_mult = 0.01
        self.position_lr_max_steps = 30_000
        self.feature_lr = 0.0025
        self.opacity_lr = 0.05
        self.scaling_lr = 0.005
        self.rotation_lr = 0.001
        self.percent_dense = 0.001
        self.lambda_dssim = 0.2
        self.densification_interval = 100
        self.opacity_reset_interval = 3000
        self.densify_from_iter = 500
        self.densify_until_iter = 15_000
        self.densify_grad_threshold = 0.0002
        
        self.single_view_weight = 0.015
        self.single_view_weight_from_iter = 7000  # 7000

        self.use_virtul_cam = False
        self.virtul_cam_prob = 0.5
        self.use_multi_view_trim = True
        self.multi_view_ncc_weight = 0.15
        self.multi_view_geo_weight = 0.03
        self.multi_view_weight_from_iter = 7000  # 7000
        self.multi_view_patch_size = 3
        self.multi_view_sample_num = 102400
        self.multi_view_pixel_noise_th = 1.0

        self.opacity_cull_threshold = 0.005
        self.densify_abs_grad_threshold = 0.0008
        self.abs_split_radii2D_threshold = 20
        self.max_abs_split_points = 50_000
        self.max_all_points = 6000_000
        self.exposure_compensation = False
        self.random_background = False
        super().__init__(parser, "Optimization Parameters")

[jaco001]

If you use in train.py "-r 1" parameter for my side it always do a random thing. If I set it like -r 2400 it ten to work.

[fy99925]

Oh, I misunderstood. Okay, I'll try that. I did use -r1 to train.

[fy99925]

Well, I used the -r 5605(the width of images) parameter for the training, and the results were about the same as the -r 1, both more blurred than the training results automatically cropped to 1.6k.

This is confusing for me, but thank you for your help.

[AsherJingkongChen]

The key is your image count, not size.

[NotCompsky]

I used -r 5605(width of images) and the results were about the same as the -r 1, both more blurred than the training results automatically cropped to 1.6k.

@fy99925 Did you eventually discover what the problem was, or have another way to get high resolution images to work well?

[BitCalSaul]

I used A100 on the server to train the original resolution image (almost 5600*3600), with all parameters unchanged and 30,000 iterations. The result is blurrier than the image training that is automatically cropped to 1.6k. It is strange. I think higher resolution images may lead to better results.

I wonder about this, what is the effect of high resolution images on guassian splatting? How can I use high resolution images better in gs?

Hi, do you both compare with the original resolution ground truth figure for PSNR metric?

[BitCalSaul]

If you use in train.py "-r 1" parameter for my side it always do a random thing. If I set it like -r 2400 it ten to work.

I think -r 1 means keeping the original resolution? hard to image there is some randomness.

[BitCalSaul]

@NotCompsky Hi do you have any findings?