Update README.md
Browse files
README.md
CHANGED
|
@@ -16,14 +16,14 @@ tags:
|
|
| 16 |
- paper-implementation
|
| 17 |
|
| 18 |
results:
|
| 19 |
-
Metric
|
| 20 |
-
mAP@0.5
|
| 21 |
-
RPN Recall @ IoU 0.5
|
| 22 |
|
| 23 |
Known_Deviations(expected affect on mAP):
|
| 24 |
-
-ResNet-50 instead of ResNet-101
|
| 25 |
-
-Batch size 2 instead of 1
|
| 26 |
-
-No horizontal flip augmentation
|
| 27 |
-
-BatchNorm frozen from Step 2 onward
|
| 28 |
|
| 29 |
---
|
|
|
|
| 16 |
- paper-implementation
|
| 17 |
|
| 18 |
results:
|
| 19 |
+
- Metric Result
|
| 20 |
+
- mAP@0.5 63%
|
| 21 |
+
- RPN Recall @ IoU 0.5 - 83%
|
| 22 |
|
| 23 |
Known_Deviations(expected affect on mAP):
|
| 24 |
+
-ResNet-50 instead of ResNet-101 due to compute constraint resulting weaker features than ResNet-101, likely costing several mAP points — probably felt most on small/textured classes like `bottle`/`pottedplant`, this project's weakest. Not isolated by a ResNet-101 run.
|
| 25 |
+
-Batch size 2 instead of 1, GPU throughput, Paper's lr schedule (per-image, batch size 1) reused unscaled, changing gradient noise per step. Not isolated.
|
| 26 |
+
-No horizontal flip augmentation, not implemented, Paper's VOC recipe uses flipping as a free 2× augmentation; skipping it likely costs some mAP, more on sparser classes. Not isolated.
|
| 27 |
+
-BatchNorm frozen from Step 2 onward, Batch size 2 is too small for stable BN statistics — standard practice, not ad hoc, Expected neutral-to-beneficial vs. unfrozen (paper's VGG16 has no BN to compare against). Step 1 is the exception — its backbone trains with BN unfrozen.
|
| 28 |
|
| 29 |
---
|