machine learning - DRG - What is a difference beetwen agent mAP and role mAP

Question

---------Reporting Agent AP (%)------------------
                hold: AP = 77.70 (#pos = 3608)
               stand: AP = 89.81 (#pos = 4118)
                 sit: AP = 73.60 (#pos = 1916)
                ride: AP = 80.96 (#pos = 556)
                walk: AP = 65.01 (#pos = 597)
                look: AP = 70.53 (#pos = 3347)
                 hit: AP = 87.92 (#pos = 349)
                 eat: AP = 84.23 (#pos = 521)
                jump: AP = 58.83 (#pos = 635)
                 lay: AP = 35.05 (#pos = 387)
       talk_on_phone: AP = 84.42 (#pos = 285)
               carry: AP = 65.75 (#pos = 472)
               throw: AP = 47.57 (#pos = 244)
               catch: AP = 60.25 (#pos = 246)
                 cut: AP = 76.33 (#pos = 269)
                 run: AP = 76.46 (#pos = 687)
    work_on_computer: AP = 79.94 (#pos = 410)
                 ski: AP = 84.38 (#pos = 424)
                surf: AP = 93.78 (#pos = 486)
          skateboard: AP = 95.75 (#pos = 417)
               smile: AP = 55.91 (#pos = 1415)
               drink: AP = 31.04 (#pos = 82)
                kick: AP = 79.93 (#pos = 180)
               point: AP = 1.04 (#pos = 31)
                read: AP = 55.99 (#pos = 111)
           snowboard: AP = 87.57 (#pos = 277)
Average Agent AP = 69.22
---------------------------------------------
---------Reporting Role AP (%)------------------
               hold-obj: AP = 34.63 (#pos = 3608)
              sit-instr: AP = 29.34 (#pos = 1916)
             ride-instr: AP = 65.95 (#pos = 556)
               look-obj: AP = 36.36 (#pos = 3347)
              hit-instr: AP = 74.28 (#pos = 349)
                hit-obj: AP = 44.93 (#pos = 349)
                eat-obj: AP = 39.51 (#pos = 521)
              eat-instr: AP = 5.24 (#pos = 521)
             jump-instr: AP = 52.71 (#pos = 635)
              lay-instr: AP = 25.35 (#pos = 387)
    talk_on_phone-instr: AP = 54.31 (#pos = 285)
              carry-obj: AP = 34.82 (#pos = 472)
              throw-obj: AP = 40.64 (#pos = 244)
              catch-obj: AP = 46.16 (#pos = 246)
              cut-instr: AP = 38.49 (#pos = 269)
                cut-obj: AP = 40.95 (#pos = 269)
 work_on_computer-instr: AP = 60.67 (#pos = 410)
              ski-instr: AP = 51.41 (#pos = 424)
             surf-instr: AP = 79.32 (#pos = 486)
       skateboard-instr: AP = 86.30 (#pos = 417)
            drink-instr: AP = 18.52 (#pos = 82)
               kick-obj: AP = 66.37 (#pos = 180)
            point-instr: AP = 0.24 (#pos = 31)
               read-obj: AP = 32.88 (#pos = 111)
        snowboard-instr: AP = 72.92 (#pos = 277)
Average Role [scenario_1] AP = 45.29

Hi, I try to understand what is the difference beetwen agent mAP and role mAP. Can U explain what is the difference? Thanks in advance.

I downloaded code from:

https://github.com/vt-vl-lab/DRG/

Paper:

https://arxiv.org/pdf/2008.11714.pdf

Answer 1

等待大神答复