floating things wheen training nerf-synthetic datasets

[Moreland-cas]

Hi~I believe nerfstudio are using nerfacc to train instant-ngp models of nerf-synthetic dataset, when training on ficus dataset, I notice there are floating things in the scene cube, which hurt the psnr drastically, whereas no such things exist in scene like lego.
Here is a depth map picture rendered from ficus dataset:

any idea on how to fix this? Thanks Here is a copy of my config: ngp_config = MethodSpecification( config=TrainerConfig( method_name="ngp", steps_per_eval_batch=200, steps_per_save=2000, steps_per_eval_all_images=2000, max_num_iterations=20000, mixed_precision=False, use_grad_scaler=False, pipeline=DynamicBatchPipelineConfig( datamanager=VanillaDataManagerConfig( dataparser=BlenderDataParserConfig() ), target_num_samples = 1<<18, max_num_samples_per_ray = 2**8, model=SegNGPModelConfig(eval_num_rays_per_chunk=8192), ), optimizers={ "fields": { "optimizer": AdamOptimizerConfig(lr=1e-2, eps=1e-15, weight_decay=1e-5), # 1e-5 if in ["materials", "ficus", "drums"] else 1e-6 "scheduler": MultiStepSchedulerConfig( max_steps=20000, gamma=0.33, milestones=(10000, 15000, 18000) # need LinearLR ? ), }, }, vis="wandb", ), description="Segment NGP config", )