make repo self-contained: ship backbone.safetensors; from_pretrained prefers it

Ship imagenet_classes.json so standalone ArgusLite returns real ImageNet labels instead of class_0..999 placeholders.

README: restore depth row RMSE 0.537

depth_head: restore prior checkpoint

README: update depth row

depth_head: retrain best checkpoint

README: drop training log references

drop training-log process artifact

drop training-log process artifact

drop training-log process artifact

drop training-log process artifact

README: state segmentation head is a linear probe (BN + 1x1 conv)

Depth head: reflection-padded DPT convs (RMSE 0.537)

Depth head: reflection-padded DPT convs (RMSE 0.537)

Depth head: reflection-padded DPT convs (RMSE 0.537)

Depth head: ship improved checkpoint (RMSE 0.542)

Depth head: ship improved checkpoint (RMSE 0.542)

Update README.md

Update infer.py

Update argus_lite.py

Depth head: DPT decoder over 4 hooked blocks, RMSE 0.670 (vs 0.715 simple head)

Seg head: 32 epochs (40k iters), mIoU 0.419 plateau

Argus-Lite detection: COCO val2017 mAP 0.273 / AP@50 0.496 / AR@100 0.432

Rewrite Caveats section as Evaluation details; strip personal-context and apologetic framing

Argus-Lite classifier: val top-1 79.13% / top-5 95.53% (beats EUPE-ViT-S paper kNN 78.2)

Add RF100-VL cross-domain eval (0.266 mean AR@100, between FCOS and Argus-B)

Add classifier health-check note to caveats

Argus-Lite: classifier head + full README assembling all four heads

Detection head trained on COCO ViT-S (2 epochs, loss cls 0.25 / reg 0.33 / ctr 0.60)

Depth head trained on NYU ViT-S (3 epochs, RMSE 0.715)

Segmentation head trained on ADE20K ViT-S features (mIoU 0.417)

initial commit