Replace SurgMotion-ONNX with 5D tubelet export (README.md)

models/SurgMotion-ONNX/README.md

@@ -5,15 +5,15 @@ Compliance-facing ONNX distribution for the SurgMotion ViT-S doctor-demo encoder
 ## Model
 
 - **Product name:** SurgMotion-ONNX
-- **Architecture:** SurgMotion ViT-S/16 video backbone (`tubelet_size=2`, SDS single-frame input duplicated internally)
+- **Architecture:** SurgMotion ViT-S/16 video backbone (`tubelet_size=2`, SDS prev+current frame pair)
 - **Checkpoint source:** `/home/chen_chuxi/NSJepa_jinlin/ckpts/SurgMotion-vits.pt`
 - **Intended use:** Offline/online surgical phase feature extraction in the SurgMotion Agent demo
-- **Input:** `pixel_values` float32 `[batch, 3, 224, 224]` with ImageNet normalization
+- **Input:** `pixel_values` float32 `[batch, 3, 2, 224, 224]` with ImageNet normalization (time dim: prev, current)
 - **Primary output used by SurgMotion:** `last_hidden_state` → drop CLS token → mean pool → 384-d clip feature
 
 ## Provenance
 
-Exported from SurgMotion ViT-S encoder weights in the bundled pretraining checkpoint. External ONNX I/O matches the doctor-demo bundle (`pixel_values` single RGB frame). Internally the exporter duplicates the frame once to satisfy `tubelet_size=2` while keeping the SDS/ONNX runtime contract unchanged.
+Exported from SurgMotion ViT-S encoder weights in the bundled pretraining checkpoint. External ONNX I/O expects a 2-frame tubelet (`pixel_values` shape [B,3,2,224,224]) with the selected sample frame and its immediate previous frame.
 
 ## Files