Pavel Konovalov

@tancik With the last changes still got None, this time during training on the Lego scene with the default number of steps with a disabled collider, camera optimizer, and average...

Before I saw this on my own real scene, now on Lego. Looks like some randomness involved, on most of trials I do not see errors. Will try anomaly detection...

Thanks. I have removed the `self.collider = NearFarCollider(near_plane=self.config.near_plane, far_plane=self.config.far_plane) ` line in the nerfactor model to disable the collider for test. Results look similar without it and with --pipeline.model.near-plane 2....

@Madaoer @tancik 1. I have found that this problem also occurs for both instant_ngp and nerfacto models. 2. I have found that in some point: at https://github.com/nerfstudio-project/nerfstudio/blob/d767cd50118a1abe0167a6316fc565685eca1926/nerfstudio/fields/instant_ngp_field.py#L148 ``` density_before_activation.max() Out[26]:...

@tancik as for reproductions, I think that it is enough training nerfacto or instant ngp on Lego with a bigger number of steps, after ~ 1 hour of training. Now...

@Madaoer as an experiment you can probably also try to clip the values of base_mlp_out , for example by the `base_mlp_out .clamp(-1000, 1000)`. On my synthetic scenes I actualy see...

@krrish94 I have tried. Didn't help

@nickscamara yes, we have developed a graph based retriever that is very different from standart implementations and beat all standart retrievers with significant gap on all datasets that we have...