allow min/max size kwargs for carla multimodal frcnn

[swsuggs]

You can get better benign carla performance if you pass "min_size":960, "max_size":1280 through to the underlying FRCNN model. This works automatically for the unimodal model but requires this minor fix for multimodal.

To test/use, set the specified parameters under "model_kwargs" in the config.

For reference, here are my results without and with these kwargs, using the CARLAAdversarialPatchPyTorch attack.

(mAP, disappearance, hallucination, misclass, tpr)

Single Modal Undefended
        Benign                    Adv
w/out	65; 29; 3.1; 3; 68        12; 81; 67; 1; 18
with    78; 15; 6.2; 4; 81         3; 91; 92; 0;  9

Multimodal undefended
        Benign                    Adv
w/out   66; 29; 2.9; 3; 68        27; 48; 26; 3; 49
with    79; 13; 4.5; 4; 83        35; 32; 29; 3; 65

Multimodal defended
        Benign                    Adv
w/out   70; 28; 1.9; 3; 69        35; 46; 8.7; 3; 51
with    80; 14; 2.8; 3; 83        40; 33; 12.3; 3; 64

[deprit]

Everything looks good.

[swsuggs]

For future reference, here is the result for Multi-object tracking. The difference is reduced.

(assa, deta, hota)

Advpatch undefended
        Benign        adv
w/out   62; 49; 55    56; 15; 29
With    64; 55; 59    58; 15; 29

Dpatch undefended
        Benign        adv
w/out   62; 49; 55    59; 38; 47
With    64; 55; 59    61; 42; 50

Dpatch defended
        Benign        adv
w/out   53; 34; 42    51; 24; 34
With    53; 36; 44    49; 25; 35