[TODO]: Add nuScene metric computation

[KSeangTan]

Description

It seems that the AP computation is slightly different from the AP computation from nuScenes, where it computes and interpolates a precision for 101 recall bins (0.0 - 1.0) to smooth the PR curve. However, it simply gets the maximum precision for a recall bin, which I believe it is similar to 11-point interpolated AP

I am suggesting that we should also add a feature to filter out min_recall/min_precision, and also TP errors, for example, trans_error, orientation error, therefore, it can decreases the discrepancy if someone evaluates models based on nuScenes metrics.

Purpose

1. To reduce discrepancy for models that evaluate in different metrics

Possible approaches

1. Add a class NusceneAP for the 101 recall bins, and make Map to support it

Definition of done

1. Implement the nuScene metrics as described

[KSeangTan]

@ktro2828 @scepter914

Would you mind sharing your feedback about it? Thanks

[KSeangTan]

In addition, I believe it is also important if we would like to apply this repo into our internal repo

[scepter914]

@KSeangTan
Sorry for reply.
I agreed with it.

@ktro2828 How about you?

[ktro2828]

@KSeangTan Thank you for proposing!
Looks good to me, but I have one question about who tackles this task.

[KSeangTan]

Hi @scepter914 @ktro2828
Thanks for confirming this, I will work on this since it's part of evaluation pipeline in AWML