Update README.md
Browse files
README.md
CHANGED
|
@@ -60,8 +60,6 @@ checkpoint = torch.load(ckpt_path, map_location="cpu")
|
|
| 60 |
# in the linked GitHub repo for the full model class definitions and inference pipeline.
|
| 61 |
```
|
| 62 |
|
| 63 |
-
Full inference pipeline (detection, tracking, embedding, ranking, and a Gradio demo UI):
|
| 64 |
-
[GitHub repo link — add yours here]
|
| 65 |
|
| 66 |
## License
|
| 67 |
|
|
@@ -84,5 +82,4 @@ separately, as they are not addressed here.
|
|
| 84 |
performance on unrelated camera angles/lighting is untested.
|
| 85 |
- During the contrastive stage, the visual encoder's ResNet-50 backbone was fine-tuned
|
| 86 |
further (unfrozen), while the CLIP text backbone was kept frozen and only its projection
|
| 87 |
-
head was trained
|
| 88 |
-
experiment with unfreezing the text backbone too.
|
|
|
|
| 60 |
# in the linked GitHub repo for the full model class definitions and inference pipeline.
|
| 61 |
```
|
| 62 |
|
|
|
|
|
|
|
| 63 |
|
| 64 |
## License
|
| 65 |
|
|
|
|
| 82 |
performance on unrelated camera angles/lighting is untested.
|
| 83 |
- During the contrastive stage, the visual encoder's ResNet-50 backbone was fine-tuned
|
| 84 |
further (unfrozen), while the CLIP text backbone was kept frozen and only its projection
|
| 85 |
+
head was trained.
|
|
|