Open ylqi007 opened 5 years ago
Hello, I would like to know whether the code you used is balancap/ ssd-tensorflow. I also did the same repetition, but the effect was very poor and the mAP was very low.I would like to ask you, what changes have you made to the code?thank you @ylqi007
@fuaiguo The code is from balancap/ssd-tensorflow. The training script is shown as below:
CHECKPOINT_PATH=./checkpoints/ssd_300_vgg.ckpt
DATASET_DIR=/tmp/tfrecords_pascal_2007
TRAIN_DIR=./logs/PASCAL_VOC_2007/
python3 train_ssd_network.py \
--train_dir=${TRAIN_DIR} \
--dataset_dir=${DATASET_DIR} \
--dataset_name=pascalvoc_2007 \
--num_classes=21 \
--dataset_split_name=train \
--model_name=ssd_300_vgg \
--save_summaries_secs=60 \
--save_interval_secs=600 \
--checkpoint_path=${CHECKPOINT_PATH} \
--optimizer=adam \
--max_number_of_steps=7000 \
--learning_rate=0.001 \
--batch_size=32 \
--match_threshold=0.5
By the way, have you modified the original code? I think there are some differences between the code and the paper. According to the original code, the effect is very poor.
@fuaiguo No. I used the original code.
I also got a really low mAP AP_VOC07/mAP[1.7213194274951696e-05] AP_VOC12/mAP[1.1840413668554323e-05] Could anyone please help me with this!!!
I also got a really low mAP
Could anyone please help me with this!!!
CHECKPOINT_PATH=./checkpoints/ssd_300_vgg.ckpt checkpoint_path=./checkpoints/vgg_16.ckpt This is the essence.
I have two questions about the evaluation, could anyone give me some hint and help, please?
AP_VOC07 v.s. AP_VOC12 When I ran the
eval_ssh_network.py
, I got two results,AP_VOC07/mAP
andAP_VOC12/mAP
. But in the original paper and this repo, only one result is shown. Which kind of result is shown in the original paper and this repo,AP_VOC07/mAP
orAP_VOC12/mAP
?I didn't get the result like the author shown in this repo. When I finetuned on
VOC07 train
dataset based onssd_300_vgg.ckpt
, and evaluated onVOC07 test
dataset, I got the result like bellow, and theAP_VOC07/mAP
matches the result shown in the SSD paper, But when I finetuned onVOC07+12 trainval
dataset based onssd_300_vgg.ckpt
, and evaluated onVOC07 test
dataset, I got the result like bellow, and the results, bothAP_VOC07
andAP_VOC12
, are different from the result (0.778) shown in this repo,This is the result from this repo
I am confusing now. Could anyone give me some hint and help?
Sincerely