Cannot load NeRF dataset with more than 42 images

[snaiperskaya96]

As per title, I've got a dataset with 70+ images. When trying to open it, instant-ngp progress bar freezes at the last 2 images, something like this

19:41:34 SUCCESS Initialized CUDA 11.5. Active GPU is #0: NVIDIA GeForce RTX 4090 [89]
19:41:34 INFO Loading NeRF dataset from
19:41:34 INFO .\transforms.json
19:41:34 PROGRESS [==========================================================================================================================================> ] 95% (41/43) 0s/0s

I noticed this is not a problem as far as the number of images stays under 43, is this a known limitation and is there a workaround if so?
I've tried with both, latest binary release and a manually compiled build on the master branch on this commit.

I'm sorry if I'm not seeing anything obvious, this is not really my kinda thing and I'm somewhat new to it.

Cheers,
Jury

[snaiperskaya96]

Sorry - i was a little too fast at opening this issue.
After fiddling a little with the transforms.json file i've figured those two transforms are causing the issue above, after removing them it loads up just fine. Not sure if it can be of any help but this is what they look like:

  {
    "file_path": "./images/IMG_20250525_114347_00_044.jpg",
    "sharpness": 20.799952468616283,
    "transform_matrix": [
      [
        0.9203493089370172,
        -0.011337391643574859,
        -0.39093300332656056,
        -1.5062260814896753
      ],
      [
        -0.3909221556277797,
        0.0032494568249035664,
        -0.9204180079016666,
        -1.2194180389325184
      ],
      [
        0.011705459347119722,
        0.9999304498719466,
        -0.001441402237281989,
        0.002175181118590503
      ],
      [
        0,
        0,
        0,
        1
      ]
    ]
  },
  {
    "file_path": "./images/IMG_20250525_114219_00_042_8_2.jpg",
    "sharpness": 18.036667781313266,
    "transform_matrix": [
      [
        0.055859890049010076,
        -0.0015532160731813508,
        0.998437409256881,
        -1.5196311522483346
      ],
      [
        0.9984226645766474,
        -0.005565994249523759,
        -0.055867723845543345,
        -4.097059639613031
      ],
      [
        0.005644071525082155,
        0.9999833034745602,
        0.0012398503036571823,
        0.0378853110771259
      ],
      [
        0,
        0,
        0,
        1
      ]
    ]
  }

I'm also attaching the whole file (without the entries above) in case you want to take a look.
transforms.json

[snaiperskaya96]

One more (and perhaps last) thing i've noticed is that those two pictures referred by the transforms above are partially corrupt, i cannot share the whole thing unfortunately but they are normal pictures with a chunk of filled pixels on the bottom, probably due to some issue with the panorama splitter (i've used aliceVision_split360Images).

Here's a small section of the picture with the corrupt chunk of pixels

Perhaps the takeaway from this whole thing is that instant-ngp should report whenever it cannot load the image or the image does not make sense, not sure which one applies here.