tobil/omatrack-telemetry-models

1{2  "schema": "gauge-detector-center-v1",3  "training_scope_version": "pilot-v2-mil-strong4-backgrounds",4  "tensor_decode_abi": "unchanged from gauge-detector-center-v1",5  "status": "experimental_discovery_not_reader_approval",6  "input": {"name":"image","dtype":"float32","layout":"NCHW","shape":[1,3,384,640],"channels":"RGB","range":[0,1],"spatial_multiple":32,"dynamic_spatial":true,"resize":"stretch_full_source_half_pixel_bilinear_no_antialias","pixel_formula":"sx=(dx+0.5)*source_width/dest_width-0.5; sy analogous; floor coordinates, clamp each neighbor index; interpolate raw uint8 RGB in float32 then divide by 255; no intermediate uint8 rounding","normalization":"divide_255_only; ImageNet normalization is inside the graph"},7  "stride":4,8  "outputs":[9    {"name":"center_logits","channels":1},10    {"name":"size_logits","channels":2,"order":["width","height"]},11    {"name":"offset_logits","channels":2,"order":["x","y"]},12    {"name":"representation_logits","channels":4,"labels":["digits","bar","wheel","needle"]},13    {"name":"semantic_logits","channels":5,"labels":["unknown","gear","stint_lap","brake","throttle"]}14  ],15  "output_layout":"NCHW; all outputs [1,C,input_height/4,input_width/4]; finite float32 logits",16  "decode":{"score_threshold":0.35,"local_max_kernel":3,"local_max_rule":"sigmoid(center) >= every in-bounds 3x3 neighbor; equal maxima all retained before top-k","candidate_limit":100,"sort":"descending score then ascending flat y*grid_width+x index","center":"((x+sigmoid(offset_x))/grid_width,(y+sigmoid(offset_y))/grid_height)","size":"(sigmoid(width_logit),sigmoid(height_logit)) normalized full-source width/height","box":"center +/- size/2; clip each coordinate to [0,1]; discard zero-area","nms":"class-agnostic greedy suppress IoU strictly > 0.4","nms_iou":0.4,"max_detections":32,"representation":"argmax; lowest index wins ties","semantic":"softmax argmax; lowest index wins ties; emit unknown if max probability < 0.8","semantic_threshold":0.8},17  "runtime_guards":["require RGB source dimensions > 0","require batch=1,3 input channels, H/W divisible by32","reject missing, extra, wrong-shaped, nonfloat32 or nonfinite outputs","model and metadata SHA256 must match pinned setup identity"],18  "limitations":["scores are not calibrated probabilities","unknown semantic has positive supervision only for explicitly observed non-Tier1 fields in four TRAIN images; ambiguous fields and generic lap-counter ownership remain masked; no open-set certification","all synthetic semantic loss fully masked; no renderer-intent labels","real weak windows use positive MIL only, no extent or offset loss or background negatives; four pixel-reviewed TRAIN images provide stronger geometry with explicit ignored regions","bar includes segmented indicators, no fill direction or scale predicted","visibility/readability/support remain unknown","detector boxes never imply arbitrary crop reader support","confirmed user/reviewed configuration required before extraction","only two known real layouts and four reviewed TRAIN background assets; repeated views are not new source diversity"]19}20