Upload competitors/reported_gaze.json with huggingface_hub

competitors/reported_gaze.json

@@ -1,8 +1,8 @@
 {
   "_comment": "Gaze accuracy as REPORTED by each method (within-dataset / in-distribution, the standard normalized protocol), for context vs our held-out cross-tool harness. py-feat is from the au_deep v24_s3 run that produced the SHIPPED v2.4 checkpoint; OF3 and L2CS from their papers; LibreFace from its landmark-gaze paper. mean angular error degrees (LibreFace's is per-axis-MAE-derived, landmark-based). LibreFace 2.0 DOES report gaze (a MediaPipe-landmark MLP), contrary to an earlier note here.",
   "reported": [
-    {"tool": "pyfeat_v2", "dataset": "mpiigaze", "value": 4.24, "source": "au_deep v2.4 (shipped ckpt, v24_s3 run; self-reported, in-distribution)"},
-    {"tool": "pyfeat_v2", "dataset": "gaze360", "value": 9.40, "source": "au_deep v2.4 (shipped ckpt, v24_s3 run; self-reported, in-distribution; deployed ckpt stores gaze_mae 9.9)"},
+    {"tool": "pyfeat_v2", "dataset": "mpiigaze", "value": 3.92, "source": "au_deep v2.4 DEPLOYED ckpt (v24fix_s3, weight-verified 50/50 match to HF face_multitask_v2.pt); in-distribution"},
+    {"tool": "pyfeat_v2", "dataset": "gaze360", "value": 6.81, "source": "au_deep v2.4 DEPLOYED ckpt (v24fix_s3, weight-verified); in-distribution"},
     {"tool": "openface3", "dataset": "mpiifacegaze", "value": 2.56, "source": "OpenFace 3.0 paper (arXiv:2506.02891)"},
     {"tool": "openface3", "dataset": "gaze360", "value": 10.6, "source": "OpenFace 3.0 paper (arXiv:2506.02891)"},
     {"tool": "l2cs_net", "dataset": "mpiigaze", "value": 3.92, "source": "L2CS-Net paper (arXiv:2203.03339); py-feat's gaze lineage"},