From 0c379c3f7288f0537152bf8629bf5ecbe1fe9662 Mon Sep 17 00:00:00 2001 From: Ruozhu <ruozhu.li@wur.nl> Date: Thu, 24 Apr 2025 11:55:42 +0200 Subject: [PATCH] revise readme file --- README.md | 33 +++++++++++++++------------------ ~$experiment_results.xlsx | Bin 0 -> 165 bytes 2 files changed, 15 insertions(+), 18 deletions(-) create mode 100644 ~$experiment_results.xlsx diff --git a/README.md b/README.md index 3ba1c03..bc60163 100644 --- a/README.md +++ b/README.md @@ -77,28 +77,25 @@ See `experiment_results.xlsx` for:  ## Summary Table -| index | type | backbone | scale count(s) | augmentation | training time of 20 epochs | mAP on test set | -|--------------|-----------|-----------|----------------|----------------------------|----------------------------|------------------| -| experiment_1 | one-step | ResNet-18 | 1 | none | 2m 38s | 0.1068 | -| experiment_2 | one-step | ResNet-18 | 2 | none | 2m 43s | 0.4435 | -| experiment_3 | one-step | ResNet-34 | 2 | none | 2m 33s | 0.4595 | -| experiment_4 | one-step | ResNet-50 | 2 | none | 2m 60s | 0.4348 | -| experiment_5 | one-step | ResNet-34 | 3 | none | 2m 37s | 0.4215 | -| experiment_6 | one-step | ResNet-34 | 2 | aggressive, *3 training data| 7m 32s | 0.5455 | -| experiment_7 | two-step | ResNet-34 | 2 | aggressive, *3 training data| detector: 7m 29s<br>classifier: 6m 16s | 0.4461 | +| Index | Type | Training Time of 20 Epochs | Performance: mAP on Validation Set | Performance: mAP on Test Set | +|-------|------|----------------------------|-----------------------------------|------------------------------| +| experiment_1 | one-step | 2m 38s | 0.198 | 0.107 | +| experiment_2 | one-step | 2m 43s | 0.637 | 0.444 | +| experiment_3 | one-step | 2m 33s | 0.541 | 0.460 | +| experiment_4 | one-step | 2m 60s | 0.690 | 0.435 | +| experiment_5 | one-step | 2m 37s | 0.717 | 0.422 | +| experiment_6 | one-step | 7m 32s | 0.696 | 0.545 | +| experiment_7 | two-step | detector: 7m 29s<br>classifier: 6m 16s | 0.801 | 0.446 | ## Qualitative Results  ## Key Insights -* Multi-scale detection significantly improves performance on mixed-size fruits. -* ResNet-34 has a moderate depth and number of parameters, offering a good balance for small datasets. It extracts features better than ResNet-18 but is less prone to overfitting than ResNet-50. -* Aggressive color jittering helps models generalize better across challenging lighting and backgrounds, requiring more time but resulting in a higher mAP. -* Two-step detection is more modular but does not significantly outperform optimized one-step models in this task. +* Multi-scale detection significantly improves performance on fruits with size variants. +* ResNet-34 offers a good balance for our datasets. It extracts features better than ResNet-18 but is less prone to overfitting than ResNet-50. +* Color jittering helps models perform better across challenging backgrounds. +* Two-step method does not significantly outperform optimized one-step models in this task. -## Link to Report - -Full report: //TODO - -Code Repository: https://git.wur.nl/wei044/deeplearning-mbe-8 \ No newline at end of file +## Code Repository + https://git.wur.nl/wei044/deeplearning-mbe-8 \ No newline at end of file diff --git a/~$experiment_results.xlsx b/~$experiment_results.xlsx new file mode 100644 index 0000000000000000000000000000000000000000..80deda320a2b5c0ec0700e4eb06769f7deae5d94 GIT binary patch literal 165 icmZPzD$TFTC{-X4urUNNlrrQqR54^QlrkvL7y<zHLlGwc literal 0 HcmV?d00001 -- GitLab