Problems of KITTI360 annotations

[anslt]

The GT weight map of KITTI360 is incorrect.

[nikhilgosala]

Hey! Could you please elaborate on why you think it is incorrect?

[anslt]

You can check the weight matrix.

the weight column is the average weight for certain semantic categories by your weight matrix, the new weight column is the average weight for certain semantic categories calculated by myself based on your formula in your paper.

[nikhilgosala]

Where did you get the second weight matrix column from?

[anslt]

weight column from your dataset kitti360_panopticbev/class_weights. new weight column, calculate base on the way to generate weight in your paper

[nikhilgosala]

I have a couple of questions:

1. When you say you compute the weights based on my paper, what do you do? Do these values represent only the frequency-based weighting, or do you also include the distance-based weighting and the sensitivity-based weighting? I believe you should get the same results when you use only the distance-based and frequency-based weighting formulae.

2. Are you also sure that the order is not shuffled? Are the classes following the same structure in my weights and yours? The numbers look awefully similar but they just seem to be shuffled.

[anslt]

1. only for frequency-based
2. the order is shuffled, but you can check your value based on your order.

[nikhilgosala]

1. Yes, then you'll have an issue. The weight masks in the class_weights folder include the distance-based weighting as well. So they will be slightly different around instances where we gradually decay the weights.
2. I don't understand what you mean. Are the numbers representing the weights for the same classes in column 1 and column 2? For instance, building in column1 has almost the exact same weight as truck in column2. Wall in column1 has the same value as building in column2. This can pretty much be said for all classes. That's why I want to ensure that you've not accidentally shuffled the classes when reporting the weights.