Cannot reproduce the result

[MRRRKING]

When I use the strategy (a) to train, the performance is low. It seems that the (a) cannot produce a good model.

(1) I trained my model using the strategy (b) and got the results as below.

Car [email protected], 0.70, 0.70: bbox AP:97.2627, 89.3979, 89.0140 bev AP:89.1479, 87.2914, 86.2751 3d AP:86.4445, 77.4183, 76.9191 aos AP:97.24, 89.25, 88.79

I used the default configure. (RCNN.ROI_SAMPLE_JIT=True)

(2) I fixed some bugs when setting RCNN.ROI_SAMPLE_JIT=False. And then I got the following results:

Car [email protected], 0.70, 0.70: bbox AP:96.2758, 89.2693, 88.8836 bev AP:89.0634, 87.2874, 86.5963 3d AP:86.1678, 77.1799, 76.7783 aos AP:96.26, 89.13, 88.66

Its performance is lower than before( RCNN.ROI_SAMPLE_JIT=True ).

But I still can't reproduce the results of the pretrained model. Does anyone have suggestions? THX.

[saurabverma]

Just a coupe of questions:

1. are the (random) initial weights same in the two cases? This might explain the difference in reproducing the results, I guess.

2. how did you make RCNN.ROI_SAMPLE_JIT=False work?

Thanks

[York1996OutLook]

how much hardware and time do you use?