Does the WD-Objects dataset will be released?

[daijun10086]

Hi!
Thanks for you great work and I really enjoy it. May I ask when the full code will be released and if the WD-Objects dataset will be released or not?

[huahuo359]

Hi, thanks so much for your interest in our work.

The full code release may be delayed for a while, as I am currently occupied with some new ongoing work.

That said, I am happy to briefly explain how we modify the base code to support our method. Our implementation is based on the original PAC-NeRF MPM simulator, with an additional grid-based force field introduced to incorporate the wind force computed by LBM. We first add a new Taichi field in the MPM simulator to store the external force on the background grid:

self.force_field = ti.Vector.field(3, dtype=ti.f32, shape=(128, 128, 128), needs_grad=True)

The wind force computed by LBM is then loaded into this field through:

@ti.kernel
def from_torch_force_field(self, force_field: ti.types.ndarray()):
    for i, j, k in ti.ndrange(128, 128, 128):
        self.force_field[i, j, k] = ti.Vector([
            force_field[i, j, k, 0],
            force_field[i, j, k, 1],
            force_field[i, j, k, 2]
        ])

Then we append an additional grid operator through pre_p2g_operations, which applies the loaded force field during the grid update:

def add_force_on_grid(self):
    @ti.func
    def apply_force(I, v, impulse_modifier):
        impulse = self.force_field[I] / self.grid_m[I]
        v = v + impulse * self.dt[None]
        return v

    self.pre_p2g_operations.append(apply_force)

This operator is called in the original grid_op stage:

@ti.kernel
def grid_op(self, s: ti.i32):
    for I in ti.grouped(self.grid_m):
        if self.grid_m[I] > 1e-8 * self.dx[None] ** 3:
            v_out = self.grid_v_in[I] / self.grid_m[I] + self.dt[None] * self.gravity[None]
            for i in ti.static(range(len(self.analytic_collision))):
                v_out = self.analytic_collision[i](I, v_out)

            for j in ti.static(range(len(self.pre_p2g_operations))):
                v_out = self.pre_p2g_operations[j](I, v_out, self.impulse_modifier)

            self.grid_v_out[I] = v_out

In this way, the LBM-computed wind force is injected into the MPM simulation as an additional grid-based external force, while the original simulator structure is largely preserved.

You can refer to these code snippets for the implementation, and you may also build more of your own data:

1. MPM
2. LBM
3. GIC
4. PhysGaussian

[daijun10086]

Thanks for sharing your implementation thoughts! Btw, does the WD-Objects dataset will be released or not?

[huahuo359]

Thank you for your interest. We plan to release both the WD-Objects dataset and the simulation code in the future. More importantly, the simulation pipeline will allow users to create their own WD-Objects-style data for different scenarios and needs, which we believe is particularly valuable.

[daijun10086]

Thanks! I believe that will be valuable!