input size to clip_vehicleid.pt model

[lochappy]

Search before asking

• I have searched the BoxMOT issues and found no similar bug report.

Question

I have checked the ViT-CLIP-ReID VehicleID model here (https://github.com/Syliz517/CLIP-ReID), which have the same download link to your clip_vehicleid.pt. The original test size is 256x256, yours is only 256x128. Still dont know why that is?

[mikel-brostrom]

This is actually a bug! As you can see here. The default resizing is (256, 128):

boxmot/boxmot/appearance/backends/base_backend.py

Lines 35 to 42 in bb0f996

	# Determine input shape, depending on model name
	if "lmbn" in self.model_name:
	input_shape = (384, 128)
	elif "hacnn" in self.model_name:
	input_shape = (160, 64)
	else:
	input_shape = (256, 128)
	self.input_shape = input_shape

[mikel-brostrom]

Fixing! #2160

[mikel-brostrom]

Merged! Thanks for reporting @lochappy !

[lochappy]

1. I think "self.model_name" should be "self.weights.name", cuz "self.model_name" returned "clip" only when I used clip-based models.

2. And I think u need to update here as well:

boxmot/boxmot/appearance/reid/export.py

Lines 49 to 50 in c8335df

	if "lmbn" in str(args.weights):
	args.imgsz = [384, 128]

3. And also need to update the cfg.INPUT.SIZE_TRAIN and cfg.INPUT.SIZE_TEST accordingly here:

boxmot/boxmot/appearance/reid/registry.py

Lines 92 to 97 in c8335df

	if "clip" in name:
	from boxmot.appearance.backbones.clip.config.defaults import _C as cfg
	return MODEL_FACTORY[name](
	cfg, num_class=num_classes, camera_num=2, view_num=1
	)

4. I also noticed that those CLIP-ReID models used mean=[0.5, 0.5, 0.5] and std=[0.5, 0.5, 0.5] instead of [0.485, 0.456, 0.406] and [0.229, 0.224, 0.225]

[mikel-brostrom]

Where do you see this:

I also noticed that those CLIP-ReID models used mean=[0.5, 0.5, 0.5] and std=[0.5, 0.5, 0.5] instead of [0.485, 0.456, 0.406] and [0.229, 0.224, 0.225]

?

Both in the defaults_base.py and defaults.py configs I see:

# Values to be used for image normalization
_C.INPUT.PIXEL_MEAN = [0.485, 0.456, 0.406]
# Values to be used for image normalization
_C.INPUT.PIXEL_STD = [0.229, 0.224, 0.225]

[lochappy]

In the test log of ViT-CLIP-ReID here (https://github.com/Syliz517/CLIP-ReID):

• market_1501: https://drive.google.com/file/d/1SKtpls1rtcuC-Xul-uVhEOtFKf8a1zDt/view

• vehicleid: https://drive.google.com/file/d/110l_8I2LQ3OfZP1xElF2Jl4lRvvhweYf/view

[mikel-brostrom]

• Input to CLIP model now set by configuring from boxmot.appearance.backbones.clip.config.defaults import _C as cfg
• mean=[0.5, 0.5, 0.5] and std=[0.5, 0.5, 0.5] instead of [0.485, 0.456, 0.406] and [0.229, 0.224, 0.225] is now used for clip model
• crops resize to [256, 256] only if models are trained on veri or vehicleid, [256,128]otherwise

[mikel-brostrom]

Anything I am missing @lochappy?

[mikel-brostrom]

You can check the changes here: #2163

[lochappy]

Last point, can we replace those lines:

boxmot/boxmot/appearance/reid/export.py

Lines 37 to 43 in ddb107f

	model_name = ReIDModelRegistry.get_model_name(args.weights)
	nr_classes = ReIDModelRegistry.get_nr_classes(args.weights)
	pretrained = not (args.weights and args.weights.is_file() and args.weights.suffix == ".pt")
	model = ReIDModelRegistry.build_model(
	model_name, args.weights, num_classes=nr_classes, pretrained=pretrained, use_gpu=args.device
	).to(args.device)
	ReIDModelRegistry.load_pretrained_weights(model, args.weights)

by one line:
model = auto_backend.model.model

only? cuz i think the ReidAutoBackend does the weights loading already in its constructor.