readme.md updated

README.md

@@ -52,16 +52,38 @@ loss = DinoSpatialLoss(student_feat, teacher_feat)
 
 ---
 
-## 📈 Performance: Latent Quality vs Supervised YOLO
+## 📈 Performance: Latent Quality & Downstream Evaluation
 
-Despite not using any labels, `core-dino` shows **higher latent alignment and semantic grouping** than supervised YOLO encoders.
+Despite being trained without any labels, `core-dino` demonstrates strong latent alignment and generalization capability — both in visual similarity and downstream tasks.
 
-📊 Embedding similarity visualizations show:
-- better clustering of semantically similar regions  
-- stronger consistency across resolution changes
+### 🔍 1. Embedding Similarity (Unsupervised)
+
+Visual inspection of patch-level embedding similarity shows that `core-dino`:
+- captures finer semantic structure than supervised YOLO
+- is robust across real-world resolution shifts
 
 ![Embedding Comparison](assets/embed-sim-vs-yolo.png)
 
+---
+
+### 🛣️ 2. Downstream: Road Extraction (DeepGlobe Dataset)
+
+We evaluated `core-dino` on the [DeepGlobe Road Extraction Dataset](https://competitions.codalab.org/competitions/18467#learn_the_details), using it as a frozen backbone in a simple segmentation pipeline.
+
+- **Setup:**  
+  - Both `core-dino` and YOLO backbones were **frozen**
+  - Only a **2-layer convolutional head** was trained
+  - Task: Binary road segmentation (IoU loss)
+
+- **Result:**  
+  - `core-dino` consistently outperformed the supervised YOLO backbone across all epochs
+  - Shows superior feature quality in transfer settings
+
+<p align="center">
+  <img src="assets/downstream-deepglobe-roads.png" alt="Downstream Performance" width="600">
+</p>
+
+
 ---
 
 ## 🗂️ Model Details