Accuracy drop with Resnet34 + SSD

We are trying custom object detection with Resnet34 + SSD. The results are very poor when comparative model trained with Resnet18 + SSD combination.

Do we need to take any special consideration while training with Resnet34 + SSD?


Could you please describe more details about the result of “Resnet34 + SSD”? Any training log for it?
More, can you provide more information about comparative model trained with Resnet18 + SSD combination.

Thanks for quick response.

We are training same dataset with Restnet18+SSD and Resnet34+SSD. PFA attached log files and config files.
resnet18-training.txt (18.9 KB) resnet34-training.txt (18.9 KB) ssd_training_log_resnet18.csv (3.7 KB) ssd_training_log_resnet34.csv (3.7 KB) ssd_train_resnet18_kitti.txt (1.3 KB) ssd_train_resnet34_kitti.txt (1.3 KB) ssd_retrain_resnet18_kitti.txt (1.3 KB) ssd_retrain_resnet34_kitti.txt (1.3 KB)

After increasing dataset size the accuracy improved.


Sorry for late reply. From your previous attachment resnet34-training.txt, the mAP can get 58.569% at 10th epoch, but get worse after that. So, need to finetune batch-size or max_learning_rate and trigger experiments.