Chapter 4 BBox Sigmoid Grounding Recap

This focused page compares only the three bbox grounders currently under consideration: activated dynamic sigmoid, activated source-pinned sigmoid, and raw source-pinned sigmoid with a class-CE term. It separates natural held-out boxes from traversal-command boxes, and separates natural-bbox training from traversal-command training.

Traversal boxes do not all have the same size. The traversal standardizes the requested position, not the object extent: each source keeps its own original bbox width and height while the box center is moved through the ten predefined positions. Very large boxes are clipped to the image frame near the borders, so some edge positions can also look truncated.
External metrics use decoded RepTok images, YOLO11n top detection for spatial IoU/detection rate, and ResNet-50 ImageNet top-1 for target-class preservation. Proxy and attention plots are diagnostics only.

Mechanisms

grounderinitial statesource proposalkeepclass CE
Activated dynamic sigmoidactivateddynamic source proposalsigmoid interpolationno
Activated pinned sigmoidactivatedraw source-pinned proposalsigmoid interpolationno
Raw pinned sigmoid + CErawraw source-pinned proposalsigmoid interpolationyes, 0.01

External Results

External delta grid
Grounded-minus-direct deltas. Bars are the three-seed mean; points are individual seeds; error bars are 10,000-resample paired/clustered bootstrap intervals for the seed-averaged delta.
External absolute grid
Direct versus grounded absolute values. Grounded bars show mean across seeds with seed standard deviation.

aggregate_metrics.csv | per_seed_values.csv | per_seed_deltas.csv

Traversal-Trained Development Diagnostics

Training proxy diagnostics
Held-out proxy behavior for the newly traversal-trained grounders. The class-probe panel is meaningful only for the CE run.

traversal_training_proxy.csv

Qualitative Traversal Grids

Seed 0 outputs are shown for compactness. Rows compare direct and the three grounded mechanisms for the same class and commanded traversal positions. Green boxes are the actual commanded boxes from the evaluation manifest: their centers follow the shared ten-step traversal, while their width and height come from the original source object.

340 zebra | source 695137
Traversal boxes: natural-trained grounders: 340 zebra | source 695137. Green box is the commanded box.
404 airliner | source 545305
Traversal boxes: natural-trained grounders: 404 airliner | source 545305. Green box is the commanded box.
340 zebra | source 695137
Traversal boxes: traversal-trained grounders: 340 zebra | source 695137. Green box is the commanded box.
404 airliner | source 545305
Traversal boxes: traversal-trained grounders: 404 airliner | source 545305. Green box is the commanded box.

Attention / Keep Diagnostics

Attention trajectory
Per-step keep and specialist weights from the common 500-sample traversal diagnostic for the natural-trained candidate checkpoints.

attention_trajectory.csv

Summary Table

evalgroundermetricdirectgrounded mean +/- SDdelta95% CIseed deltas
Natural boxes: natural-trained groundersActivated dynamic sigmoidYOLO IoU0.47410.4695 +/- 0.0074-0.0045[-0.0176, 0.0088]-0.0081898868;0.0057299137;-0.011174262
Natural boxes: natural-trained groundersActivated dynamic sigmoidDetection rate0.99000.9867 +/- 0.0021-0.0033[-0.0103, 0.0037]-0.00099998713;-0.003000021;-0.0059999824
Natural boxes: natural-trained groundersActivated dynamic sigmoidResNet target-class accuracy0.95300.9587 +/- 0.00480.0057[-0.0077, 0.0193]0.0080000162;-0.00099998713;0.0099999905
Natural boxes: natural-trained groundersActivated pinned sigmoidYOLO IoU0.46870.4876 +/- 0.01080.0189[0.0045, 0.0336]0.031520605;0.0050692558;0.020117402
Natural boxes: natural-trained groundersActivated pinned sigmoidDetection rate0.98800.9857 +/- 0.0041-0.0023[-0.0097, 0.0057]-0.0029999614;-0.0069999695;0.003000021
Natural boxes: natural-trained groundersActivated pinned sigmoidResNet target-class accuracy0.95500.9537 +/- 0.0050-0.0013[-0.0143, 0.0120]0.0040000081;0;-0.0079999566
Natural boxes: natural-trained groundersRaw pinned sigmoid + CEYOLO IoU0.46640.4855 +/- 0.00590.0191[0.0047, 0.0336]0.021548361;0.024865419;0.01102829
Natural boxes: natural-trained groundersRaw pinned sigmoid + CEDetection rate0.98500.9813 +/- 0.0068-0.0037[-0.0120, 0.0050]-0.00099998713;-0.013000011;0.0029999614
Natural boxes: natural-trained groundersRaw pinned sigmoid + CEResNet target-class accuracy0.94100.9277 +/- 0.0160-0.0133[-0.0277, 0.0017]-0.014999986;0.0070000291;-0.032000005
Traversal boxes: natural-trained groundersActivated dynamic sigmoidYOLO IoU0.46000.4589 +/- 0.0009-0.0011[-0.0057, 0.0033]8.2135201e-05;-0.0012937784;-0.0022078156
Traversal boxes: natural-trained groundersActivated dynamic sigmoidDetection rate0.98470.9822 +/- 0.0007-0.0025[-0.0051, 0.0001]-0.0017999411;-0.0034999847;-0.0021999478
Traversal boxes: natural-trained groundersActivated dynamic sigmoidResNet target-class accuracy0.95250.9576 +/- 0.00140.0051[0.0006, 0.0097]0.0039000511;0.0042000413;0.0070999861
Traversal boxes: natural-trained groundersActivated pinned sigmoidYOLO IoU0.46050.4762 +/- 0.00410.0157[0.0104, 0.0211]0.018555701;0.0099299252;0.018529534
Traversal boxes: natural-trained groundersActivated pinned sigmoidDetection rate0.98230.9783 +/- 0.0005-0.0040[-0.0071, -0.0009]-0.0032999516;-0.0041999221;-0.0043998957
Traversal boxes: natural-trained groundersActivated pinned sigmoidResNet target-class accuracy0.95050.9516 +/- 0.00120.0011[-0.0049, 0.0071]0.0019999743;-0.00050002337;0.0019000173
Traversal boxes: natural-trained groundersRaw pinned sigmoid + CEYOLO IoU0.45980.4812 +/- 0.00760.0213[0.0162, 0.0264]0.028934926;0.0240556;0.010950446
Traversal boxes: natural-trained groundersRaw pinned sigmoid + CEDetection rate0.98430.9777 +/- 0.0032-0.0066[-0.0097, -0.0037]-0.0049999952;-0.011099994;-0.0038000345
Traversal boxes: natural-trained groundersRaw pinned sigmoid + CEResNet target-class accuracy0.94980.9297 +/- 0.0174-0.0201[-0.0261, -0.0140]-0.010999978;-0.004799962;-0.044399977
Traversal boxes: traversal-trained groundersActivated dynamic sigmoidYOLO IoU0.46020.4600 +/- 0.0014-0.0001[-0.0046, 0.0042]0.00084516406;-0.0021348894;0.0008880496
Traversal boxes: traversal-trained groundersActivated dynamic sigmoidDetection rate0.98480.9819 +/- 0.0005-0.0029[-0.0056, -0.0002]-0.0035000443;-0.0022999644;-0.0028999448
Traversal boxes: traversal-trained groundersActivated dynamic sigmoidResNet target-class accuracy0.95160.9544 +/- 0.00250.0028[-0.0015, 0.0071]0;0.0024000406;0.006000042
Traversal boxes: traversal-trained groundersActivated pinned sigmoidYOLO IoU0.45660.4757 +/- 0.00320.0191[0.0137, 0.0243]0.023628682;0.016345978;0.017363757
Traversal boxes: traversal-trained groundersActivated pinned sigmoidDetection rate0.98480.9762 +/- 0.0018-0.0086[-0.0121, -0.0052]-0.010500014;-0.0091000199;-0.0062000155
Traversal boxes: traversal-trained groundersActivated pinned sigmoidResNet target-class accuracy0.95170.9499 +/- 0.0006-0.0018[-0.0084, 0.0049]-0.0021999478;-0.00099998713;-0.0021998882
Traversal boxes: traversal-trained groundersRaw pinned sigmoid + CEYOLO IoU0.46150.4778 +/- 0.00830.0164[0.0113, 0.0215]0.0048620999;0.023947775;0.02025792
Traversal boxes: traversal-trained groundersRaw pinned sigmoid + CEDetection rate0.98330.9793 +/- 0.0010-0.0040[-0.0068, -0.0012]-0.002599895;-0.0046999454;-0.0046999454
Traversal boxes: traversal-trained groundersRaw pinned sigmoid + CEResNet target-class accuracy0.95540.9461 +/- 0.0054-0.0093[-0.0145, -0.0042]-0.003000021;-0.016300082;-0.0087000728

Proxy Table

grounderseedstep 0 proxy IoUstep 5 proxy IoUdeltaclass-probe accbest epoch
Activated dynamic sigmoid00.59830.61490.01670.000090
Activated dynamic sigmoid10.59830.61450.01620.000070
Activated dynamic sigmoid20.59830.61750.01920.000055
Activated pinned sigmoid00.59830.66360.06530.000090
Activated pinned sigmoid10.59830.66250.06420.000080
Activated pinned sigmoid20.59830.66370.06540.000090
Raw pinned sigmoid + CE00.30900.64320.33420.999710
Raw pinned sigmoid + CE10.30900.64680.33780.997190
Raw pinned sigmoid + CE20.30900.64200.33300.998295