SANet-Keras

An unofficial implementation of SANet for crowd counting in Keras-TF.

---

Paper:

• Original_paper: Cao, X., Wang, Z., Zhao, Y., & Su, F. (2018). Scale Aggregation Network for Accurate and Efficient Crowd Counting. The European Conference on Computer Vision (ECCV), 1–17.

Results now:

On dataset ShanghaiTech B

Still far from the performance in the original paper(MAE 8.4)

MAE	MSE	MAPE	Mean DM Distance
13.727	22.726	11.9	27.065

Dataset:

• ShanghaiTech dataset: dropbox or Baidu Disk.

Training Parameters:

1. Loss = ssim_loss + L2

2. Optimizer = Adam(lr=1e-4)

3. Data augmentation: Flip horizontally.

4. Patch: No patch, input the whole image, output the same shape DM.

5. Instance normalization: No IN layers at present, since network with IN layers is very hard to train and IN layers didn't show improvement to the network in my experiments.

6. Output Zeros: The density map output may fade to zeros in 95%+ random initialization, I tried the initialization method in the original paper while it didn't work. In the past, when this happens, I just restarted the kernel and re-run. But now, I tried to train different modules(1-5) separately in the first several epochs to get relatively reasonable weights:

   , and it worked out to greatly decrease the probability of the zero-output-phenomena. Any other question, welcome to contact me.

7. Weights: Got best weights in 251-th epoch(300 epochs in total), and here is the loss records:

   

8. Prediction example:

   

Run:

1. Download dataset;
2. Data generation: run thegenerate_datasets.ipynb .
3. Run the main.ipynb to train the model and do the test.

Abstraction:

1. Network = encoder + decoder, model plot is here:

   Network	encoder	decoder
   Composition	scale aggregation module	conv2dTranspose
   Usage	extract multi-scale features	generate high resolution density map

2. Loss:

   Loss =

3. Normalization layer:

   • Ease the training process;
   • Reduce 'statistic shift problem'.